Sarndal (2007)

Component of Statistics Canada Catalogue no.
12-001-X Business Survey Methods Division
Article The calibration approach in survey theory and practice

by Carl-Erik Srndal
December, 2007
SurveyMethodology,December2007 Vol.33,No.2,pp.99119 StatisticsCanada,Catalogueno.12001X
99
Thecalibrationapproachinsurveytheoryandpractice
CarlErikSrndal1 Abstract
Calibrationis the principaltheme in many recentarticles onestimationin survey sampling. Wordssuch ascalibration approach and calibration estimators are frequently used. As article authors like to point out, calibration provides a systematicwaytoincorporateauxiliaryinformationintheprocedure. Calibrationhasestablisheditselfasanimportantmethodologicalinstrumentinlargescaleproductionofstatistics.Several national statistical agencies have developed software designed to compute weights, usually calibrated to auxiliary informationavailableinadministrativeregistersandotheraccuratesources. Thispaperpresentsareviewofthecalibrationapproach,withanemphasisonprogressachievedinthepastdecadeorso. Theliteratureoncalibrationisgrowingrapidlyselectedissuesarediscussedinthispaper. Thepaperstartswithadefinitionofthecalibrationapproach.Itsimportantfeaturesarereviewed.Thecalibrationapproach is contrasted with (generalized) regression estimation, which is an alternative but conceptually different way to take auxiliaryinformationintoaccount.Thecomputationalaspectsofcalibrationarediscussed,includingmethodsforavoiding extremeweights.Intheearlysectionsofthepaper,simpleapplicationsofcalibrationareexamined:Theestimationofa populationtotalindirect,singlephasesampling.Generalizationtomorecomplexparametersandmorecomplexsampling designsarethenconsidered. Acommonfeatureofmorecomplexdesigns(samplingintwoormorephasesorstages)isthat theavailableauxiliaryinformationmayconsistofseveralcomponentsorlayers.Theusesofcalibrationinsuchcasesof compositeinformationarereviewed.Laterinthepaper,examplesaregiventoillustratehowtheresultsofthecalibration thinking may contrast with answers given by earlier established approaches. Finally, applications of calibration in the presenceofnonsamplingerrorarediscussed,inparticularmethodsfornonresponsebiasadjustment. Key Words: Auxiliary informationWeightingConsistencyDesignbasedinference Regression estimator Models NonresponseComplexsamplingdesign.
1. Introduction
1.1 Calibrationdefined It is useful in this paper to refer to a definition of the calibrationapproach.Iproposethefollowingformulation. Definition.The calibrationapproach toestimationforfinite populationsconsistsof (a) acomputationofweightsthatincorporatespecified auxiliary information and are restrained by calibrationequation(s), the use of these weights to compute linearly weighted estimates of totals and other finite populationparameters:weighttimesvariablevalue, summedoverasetofobservedunits, an objective to obtain nearly design unbiased estimates as long as nonresponse and other non samplingerrorsareabsent.
(b)
(c)
In the literature, calibration frequently refers to (a) alone I shall often use the term for (a) to (c) together. Earlierdefinitions,althoughlessextensive,agreeessentially with mine. Ardilly (2006) defines calibration (or, more precisely,calagegnralis)asamethodofreweighting usedwhenonehasaccesstoseveralvariables,qualitativeor
quantitative, on which one wishes to carry out, jointly, an adjustment. Kott (2006) defines calibration weights as a set of weights,forunitsinthesample,thatsatisfyacalibrationto known population totals, and such that the resulting estimatorisrandomizationconsistent(designconsistent),or, more rigorously, that the design bias is, under mild conditions, an asymptotically insignificant contribution to the estimators mean squared error. This is the property I callnearly designunbiased. The Quality Guidelines (fourth edition) of Statistics Canada(2003)say:Calibrationisaprocedurethancanbe used to incorporate auxiliary data. This procedure adjusts the sampling weights by multipliers known as calibration factors thatmaketheestimatesagreewithknowntotals.The resulting weights are called calibration weights or final estimationweights.Thesecalibrationweightswillgenerally resultinestimatesthataredesignconsistent,andthathavea smallervariancethantheHorvitzThompsonestimator. Part (c) of the definition merits a comment. Nothing prevents producing weights calibrated to given auxiliary informationwithoutrequiring(c).Butmostpublishedwork oncalibrationisinthespiritof(c),soitmakesgoodsenseto includeit.Whennonsamplingerrorsarepresent,biasinthe estimates is unavoidable, whether they are made by calibration or by any other method. In line with (c), I
1. CarlErikSrndal,2115ErinbrookCrescent,#44,Ottawa,Ontario,K1B4J5,Canada.Email:carl.sarndal@rogers.com.
100
Srndal:Thecalibrationapproachinsurveytheoryandpractice
consider designbased inference to be the standard in this paper.Therandomizationbasedvarianceofanestimatoris thusimportant.However,thepaperfocusesonmotivations behind (point) estimation for reasons of space, the importantquestionofvarianceestimationisnotaddressed. 1.2 Commentsarising The definition in Section 1.1 prompts some comments andreferencestoearlierliterature: (1) Calibrationasalinearweightingmethod. Calibration hasanintimatelinkto practice. The fixation onweighting methods on the part of the leading national statistical agenciesisapowerful drivingforcebehindcalibration. To assignanappropriateweighttoanobservedvariablevalue, andtosumtheweightedvariablevaluestoformappropriate aggregates, is firmly rooted procedure. It is used in statistical agencies for estimating various descriptive finite populationparameters:totals,means,andfunctionsoftotals. Weightingiseasytoexplaintousersandotherstakeholders ofthestatisticalagencies. Weighting of units by the inverse of their inclusion probabilityfoundfirmscientificbackinglongagoinpapers such as Hansen and Hurwitz (1943), Horwitz and Thompson (1952). Weighting became widely accepted. Later,poststratificationweightingachievedthesamestatus. Calibration weighting extends both of these ideas. Calibration weighting is outcome dependent the weights dependontheobservedsample. Inverse inclusion probability weights are, by definition, greater than or equal to unity. A commonly heard interpretationisthatanobservedunitrepresentsitselfanda numberofothers,notobserved.Calibratedweights,onthe otherhand,arenotnecessarilygreaterthanorequaltounity, unlessspecialcareistakeninthecomputationtoobtainthis property. Calibrationisnewasaterminsurveysamplingabout 15yearsold butnotasatechniqueforproducingweights. Those who maintain I practiced calibration long before it was called calibration have a point. The last 15 years widened the scope and the appeal of the technique. Weightingakintocalibrationhaslongbeenusedbyprivate survey institutes, for example, in connection with quota sampling, a form of nonprobability sampling outside the scopeofthispaper. Weightingofobservedvariablevalueswasanimportant topic before calibration became a popular term. Some authors derived the weights via the argument that they shoulddifferaslittleaspossiblefromtheunbiasedsampling design weights (the inverse of the inclusion probabilities). Others found the weights by recognizing that a linear regression estimator can be written as a linearly weighted sum of the observed study variable values. Terms such as
StatisticsCanada,CatalogueNo.12001X
surveysampleweightingandregressionweightingand caseweightingareused.Amongsuchearlypapersare Alexander (1987), Bankier, Rathwell and Majkowski (1992), Bethlehem and Keller (1987), Chambers (1996), Fuller, Loughin and Baker (1994), Kalton and Flores Cervantes (1998), Lematre and Dufour (1987), Srndal (1982) and Zieschang (1990). I comment later on the technique repeated weighting, promoted by the Dutch national statistical agency, CBS. The newer term calibrationconveysa morespecific messageanda more definitedirectionthantheolderweighting. (2) Calibration as a systematic way to use auxiliary information.Calibration providesasystematic wayto take auxiliary information into account. As Rueda, Martnez, Martnez and Arcos (2007) point out, in many standard settings, the calibration provides a simple and practical approach to incorporating auxiliary information into the estimation. Auxiliaryinformationwasusedtoimprovetheaccuracy ofsurveyestimateslongbeforecalibrationbecamepopular. Numerous papers were written with this goal in mind, for moreorlessspecializedsituations.Today,calibration does offer a systematic outlook on the uses of auxiliary information. For example, calibration can deal effectively withsurveyswhereauxiliaryinformationexistsatdifferent levels.Intwostagesamplinginformationmayexistforthe first stage sampling units (the clusters), and other informationforthesecondstagesamplingunits.Insurveys with nonresponse (that is, essentially all surveys), infor mation may exist at the population level (known population totals), and other information at the sample level (auxiliary variable values for all those sampled, responding and nonresponding). Calibration with compositeinformationisreviewedinSections8and9. Regressionestimation,orgeneralizedregression(GREG) estimation,competeswithcalibrationasasystematicwayto incorporatingauxiliaryinformation.Itisthereforeimportant to contrastGREGestimation(describedinSection3)with calibration estimation (described in Section 4). The two approachesaredifferent. (3) Calibrationtoachieveconsistency.Calibrationisoften described as a way to get consistent estimates. (Here consistentrefersnottorandomizationconsistentbutto consistent with known aggregates.) The calibration equationsimposeconsistencyontheweightsystem,sothat, whenappliedto theauxiliary variables,itwill confirm(be consistentwith)knownaggregatesforthosesameauxiliary variables. A desire to promote credibility in published statisticsisanoftencitedreasonfordemandingconsistency. Someusersofstatisticsdislikefindingthesamepopulation
SurveyMethodology,December2007
101 allstudyvariables,ofwhichthereareusuallymanyinlarge governmentsurveys. (5) Calibration in combination with other terms. Some authors use the word calibration in combination with other terms, to describe various directions of thought. Examples of this proliferation of terms are: Model calibration (Wu and Sitter 2001) gcalibration (Vanderhoeft, Waeytens and Museux 2000) Harmonized calibration(Webber,LatoucheandRancourt2000),Higher level calibration (Singh, Horn and Yu 1998) Regression calibration(DemnatiandRao2004)Nonlinearcalibration (Plikusas2006)Supergeneralizedcalibration(Calagesuper gnralisArdilly2006)Neuralnetworkmodelcalibration estimatorandLocalpolynomialmodelcalibrationestimator (Montanari and Ranalli 2003, 2005), Modelcalibrated pseudo empirical maximum likelihood estimator (Wu 2003), and yet others. Also, calibration plays a significant roleintheindirectsamplingmethodsproposedinLavalle (2006). In a somewhat different spirit, not reviewed here, are concepts such as calibrated imputation (Beaumont 2005a), and bias calibration (Chambers, Dorfman and Wehrly (1993), Zheng and Little (2003)). The following reviewpagesdonotgivejusticetoall theinnovationswithin the sphere of calibration, but the names alone do suggest directionsthathavebeenexplored. (6) Calibration as a new direction for thought. If calibration represents a new approach with clear differences compared with predecessors, we mustexamine such questions as: Does calibration generalize earlier theories or approaches? Does calibration give better, more satisfactory answers on questions of importance, as comparedwithearlierrecognizedapproaches?Sections4.5 and7.1inthispaperillustratehowtheanswersprovidedby calibrationcomparewith,orcontrastwith,thoseobtainedin earliermodesofreasoning. Thepracticeofsurveysamplingencountersnuisances such as nonresponse, frame deficiencies and measurement errors. It is true that imputation and reweighing for non responsearewidelypracticed,throughahostoftechniques. Buttheyaresomehowseparateissues,stillwaitingtobe more fully embedded into a comprehensive, more satisfactory theory of inference in sample surveys. Many theory papers deal with estimation for an imagined ideal survey, nonexistent in practice, where nonresponse and othernonsamplingerrorsareabsent.Thisisnotacriticism of the many excellent but idealized theory papers. The foundationsneedtobeexplored,too. Sections9and10indicatethatcalibrationcanprovidea moresystematicoutlookoninferenceinsurveyseveninthe presenceofthevariousnonsamplingerrors.Futurefruitful developmentsareexpectedinthatregard.
quantity estimated by two or more numbers that do not agree. The totals with which consistency is sought are sometimes called control totals. Controlled weights or calibrated weights suggest improved, more accurate estimation.TheFrenchtermforcalibration,calage,hasa similarconnotationofstability. Consistency through calibration has a broader implica tion than just agreement with known population auxiliary totals. Consistency can, for example, be sought with appropriately estimated totals,arising inthe current survey orinothersurveys. Consistency among tables estimated from different surveys is the motive behind repeated weighting, the techniquedevelopedattheDutchnationalstatisticalagency CBSinseveralarticles:RenssenandNieuwenbroek(1997) Nieuwenbroek, Renssen, and Hofman (2000) Renssen, Kroese,andWilleboordse(2001)KnottnerusandvanDuin (2006). The stated objective is to accommodate user demandstoproducenumericallyconsistentoutputs.Asthe lastmentionedpaperpointsout,repeatedweightingcanbe seenasanadditionalcalibrationstep foranewadjustment of already calibrated weights. The final weights realize consistencywithgivenmargins. Consistency with known or estimated totals may bring the extra benefit of improved accuracy (lower variance and/or reduced nonresponse bias). However, in some articles, especially those authored in statistical agencies, consistency for user satisfaction seems a more imperative motivationthantheprospectofincreasedaccuracy. When the primary motivation for calibration is not so muchanagreementwithotherstatisticsasrathertoreduce variance and/or nonresponse bias, then balanced weight system is a more appropriate description than consistent weightsystem,becausetheobjectiveisthentobalancethe weightstoreflecttheoutcomeofthesampling,theresponse tothesurvey,andtheinformationavailable. (4) Calibration for convenience and transparency. As Harms and Duchesne (2006) point out, The calibration approachhasgainedpopularityinrealapplicationsbecause theresultingestimatesareeasytointerpretandtomotivate, relying, as they do, on design weights and natural calibrationconstraints.Calibrationonknowntotalsstrikes the typical user as transparent and natural. Users who understand sample weighting appreciate that calibration leaves the design weights slightly modified only, while respectingthecontrols.Theunbiasednessisonlynegligibly disturbed. The simpler forms of calibration invoke no assumptions, only natural constraints. Yet another advantage is appreciated by users: In many applications, calibration gives a uniqueweighting system, applicable to
102
2. Basicconditionsfordesignbasedestimation insamplesurveys
ThissectionsetsthebackgroundforSections3to7.By basic conditions I will mean single phase probability samplingofelementsandfullresponse.Inpractice,survey conditionsarenotthatsimpleandperfect,butmanytheory papersneverthelessaddressthissituation. A probability sample s is drawn from the finite population U ={1, 2, ..., k , ..., N }. Theprobabilitysampling design generates for element k a known inclusion probability, k >0, and a corresponding sampling design weight d k =1/ k. The value y ofthestudy variableyis k recordedforall k s (completeresponse).Theobjective is to estimate a population total Y =U yk with the use of auxiliary information. The study variable y may be continuousor,asinmanygovernmentsurveys,categorical. For example, if y is dichotomous with value yk =0 or yk =1 accordingas personkisemployed orunemployed, then the parameter Y =U yk to be estimated is the populationcountofunemployedpeople.(If A U isaset of elements, I write A for A.) The basic design k unbiased estimator of Y is YHT = sd k yk, the Horwitz Thompson estimator. It is, however, inefficient when powerful auxiliary information is available for use at the estimationphase. Thegeneralnotationfortheauxiliaryvectorwillbe x . k Insomecountries,forsomesurveys,thesourcesofauxiliary data permit extensive vectors x to be built. But some k examplesofsimplevectorsare:(1) x =(1, xk), where x k k isthevalueforelementkofacontinuousauxiliaryvariable x(2)theclassification vector usedtocode membershipin one of P mutually exclusive and exhaustive groups, x = k = ( g1k , ..., g pk , ..., gPk ), sothat,for p =1, 2, ..., P, k g pk =1 ifkbelongstogroupp,and g pk =0 ifnot(3)the combinationof(1)and(2), x =( , x k ) (4)thevector k k k x that codifies two classifications stringed out sideby k side,thedimensionof x beingP +Q 1,where P and Q k aretherespectivenumberofcategories,andtheminusone istoavoidasingularmatrixinthecomputationofweights calibratedtothemargins(5)theextensionof(4)tomore than two sidebyside categorical classifications. Cases 4 and 5 are particularly important for production in national statisticalagencies. In calibration reasoning it is crucially important to specify exactly theauxiliary information. Under the basic conditions we need to distinguish two different cases relativeto x : k (i) (ii)
x is a known vector value for every k U k (completeauxiliaryinformation)
x isknown(imported)total,and x isknown U k k (observed)forevery k s
It is often the survey environment that dictates whether (i)or(ii)prevails.Case(i),completeauxiliaryinformation, occurswhen x isspecifiedinthesamplingframeforevery k k (andthusknownforevery k s ).Thisenvironment U is typical of surveys on individuals and households in Scandinavia and other North European countries equipped with high quality administrative registers that can be matched with the frame to provide a large number of potentialauxiliaryvariables.Thepopulationtotal x is U k obtainedsimplybyaddingthe x . k Case (i) gives considerable freedom in structuring the auxiliary vector x . For example, if x is a continuous k k variable value specified for every k U , then we are invited to consider x2 and other functions of x for k k 2 inclusion in x , because totals such as xk and U k U log x k arereadilycomputed.Iftherelationshiptothe studyvariableyiscurved,itmaybeaseriousomissionnot totakeintoaccountknowntotalssuchasthequadraticone orthelogarithmicone. Case(ii)prevailsinsurveyswhere(i)isnotmet,but where x is imported from an outside source U k consideredaccurateenough,andtheindividualvalue x k isavailable(observedindatacollection)forevery k s. Then x issometimescalledanindependentcontrol U k total, to mark its origin from outside the survey itself. Case(ii) is lessflexible:If x isavariable withatotal k 2 xk importedfromareliablesource,then xk may U U beunavailable,barring x2 frominclusioninto x . k k
3. Generalizedregressionestimationunder thebasicconditions
3.1 TheGREGconcept Beforeexaminingcalibration,letusconsidergeneralized regression (GREG) estimation (or just regression estimation),fortwogoodreasons: (1)GREGestimationcan also be claimed to be a systematic way to take auxiliary information into account (2) some (but not all) GREG estimators are calibration estimators, in that they can be expressedintermsofacalibratedlinearweighting. GREG estimators and calibration estimators have been extensivelystudiedinthelasttwodecades.Thetermsalone, GREG estimation and calibration estimation, reflect a clear difference in thinking. Statisticians who work in the areaareoftwotypes:ThosededicatedtoGREGthinking and those dedicated to calibration thinking. The distinction may not be completely clearcut, but it helps structuring this review paper, so I will use it. I am not venturingtosaythatthelatterthinkingismoreprevalentin national statistical agencies and the former more prevalent intheacademiccircles,butperhapsthereissuchatendency.
103
(YGREG - Y ) / N =
The GREG estimator concept evolved gradually since the mid1970s. The simple(linear)GREGisexplainedin Srndal,SwenssonandWretman(1992)athoroughreview of regression estimation is given in Fuller (2002). The centralideais that predictedyvalues y canbe produced k for all N population elements, via the fit of an assisting modelandtheuseoftheauxiliaryvectorvalues x known k forall k U . The predictedvaluesservetobuildanearly designunbiasedestimatorofthepopulationtotal Y =U yk as
k YGREG = U yk + sd k ( yk - y ) . = s d k yk + ( U yk - d k yk) s
d k Ek - Ek U
) N + O
( n -1)
(3.1)
The obvious motivation behind this construction is the prospect of a highly accurate estimate Y GREG through a close fitting assisting model that leaves small residuals yk - yk. That modeling is the corner stone of GREG thinking. Some authors use the (also justifiable) name generaldifferenceestimator fortheconstruction(3.1). Thegreatvarietyofpossibleassistingmodelsgeneratesa wide family of GREG estimators of the form (3.1). The assistingmodel,animaginedrelationshipbetweenxandy, canhavemanyforms:linear,nonlinear,generalizedlinear, mixed(modelwithsome fixed,somerandom effects),and so on. Whatever the choice, the model is assisting only eventhoughitmaybeshortoftrue,(3.1)isnearlydeign unbiasedundermildconditionsontheassistingmodeland onthesamplingdesign,sothat (YGREG - Y ) / N =O p(n -1/ 2) and (YGREG - Y ) / N = (YGREG, lin - Y ) / N +O p( n -1), where , is the statistic Y , the result of linearizing Y GREG, lin GREG unbiasedfor Y. 3.2 LinearGREG BylinearGREGImeanonethatisgeneratedbyalinear fixed effects assisting model. The predictions are yk = x Bsdq with k
B s dq =
d k qk x k x k
) ( d q x y )
s k k k k
-1
so(3.1)becomes
YGREG =
where sd k Ek is the HorvitzThompson estimator in the residuals Ek = yk - x B q with BU q =( U qk x k xk)-1 k U ( qk x yk). Hence, the designbased properties U k E (YGREG)Y and Var(YGREG) Var( sd k Ek). A close fittinglinearregression ofyonxholdsthe key toasmall variancefor Y GREG (andthisisverydifferentfromclaiming thatalinearregressionisthetrueregression). The linear GREG in Srndal, Swensson and Wretman (1992) was motivated via the linear assisting model x stating that Ex ( yk )= x and Vx ( yk ) = s2. Generalized k k leastsquaresfitgivestheestimator(3.2)with qk = 1/ s2. In k that context, an educated guess about the variation of the residuals yk - xk determinesthe q . Whenthevector x k k isfixed,themodelingeffortboilsdowntoanopinionabout the residual pattern. The choice s 2 = s2xk gives the k classical ratio estimator. If qk = x for all k U and a k constant vector , then (3.2) reduces to the cosmetic form ( U x k ) s dq . B AsBeaumontandAlavi(2004)andothershavepointed out, the linear GREG estimator is biasrobust (nearly unbiased although the assisting model falls short of correct), but it can be considerably less efficient (have larger mean squared error) than model dependent alter natives which, although biased, may have a considerably smallervariance.ThusonemayclaimthatlinearGREGis not variance robust nevertheless, it is a basic concept in designbasedsurveytheory. The specification of x should include variables (with k knownpopulationtotals)thatservedalreadyindefiningthe sampling design. Design stage information should not be relinquished at the estimation stage instead, a repeated usage is recommended. For example, in stratified simple (STSI) random sampling, the vector x in estimator (3.2) k should include, along with other available variables, the dummy coded stratum identifier, k = ( g k1, gk 2, ..., g kh , ..., gkH ), where g kh =1 ifelement k belongstostratum h,and g kh =0 ifnot h =1, ..., H . We can write the linear GREG (3.2) as a weighted samplesum, YGREG = swk yk, with
wk = d k g k g k = 1 + q x k k =
xk
) B
s dq
+ d k ( yk - x B dq). k s s
(3.2)
The q arescalefactors,chosenbythestatistician.The k standardchoiceis qk =1 forall k.Thechoiceofthe q has k some(butoftenlimited)impactontheaccuracyof Y GREG nearunbiasedness holds for any specification (barring outrageous choices) for the q . Although the model is k simple, the linear GREG (3.2) contains many estimators, considering the many possible choices of the auxiliary vector x and the scale factors q . Under general k k conditions,
x k - s d k xk
) ( d q x x )
s k k k k
-1
. (3.3)
The weights w happen to becalibrated to (consistent k with)theknownpopulationxtotal: s wk x k =U xk. That Y GREG is expressible as a linearly weighted sum with calibratedweightsisafortuitousbyproduct.Itisnotpartof GREG thinking, whose central idea formulated in (3.1) is thefitofanassistingmodel.AfewotherGREGsthanthe
104
simplelinearonealsohavethecalibrationproperty,aswill benotedlater. 3.3 NonlinearGREG Two features of the linear GREG (3.2) make it a favourite choice for routine production in statistical agencies:(i)theauxiliary populationtotal x becomes U k factored out, so the estimation can proceed as long as an accurate valuefor that total canbe computed orimported, and (ii) when written as the linearly weighted sum YGREG = s wk yk, theweightsystem(3.3)isindependentof theyvariableandcan therebybeapplied toallyvariables in the survey. We need not know x individually for all k k knowing x suffices.Needlesstosay,ifwedo U U k know all x , more efficient (still nearly design unbiased) k membersoftheGREGfamily(3.1)canbesought.Thiswill also counteranothercriticism ofthelinearGREG,namely thatalinearmodelisunrealisticforsometypesofdata.For example, for a dichotomous yvariable, a logistic assisting modelmaybebothmorerealisticandyieldamoreprecise GREGestimator. By a nonlinear GREG estimator I mean one generated asin(3.1)byanassistingmodelofothertypethanlinearin x withfixedeffects.AmongthefirsttoextendtheGREG k concept in this direction are Firth and Bennett (1998) and Lehtonen and Veijanen (1998) see also Chambers et al. (1993). In the last few years, several authors have studied modelassistednonlinearGREGs. Nonlinear GREG is a versatile idea a variety of estimators become possible via assisting models x of the followingtype:
Ex ( yk | x ) = m k for k U k
not. For example, in a Labour Force Survey with I = 3 categories, employed, not employed and not in the labour force, an objective is to estimate the respective populationcounts Yi = U yik, i =1, 2, 3. Theseauthorsuse thelogisticassistingmodel
I Ex ( yik |x k ) = mik mik = exp(x i ) 1 + exp( x ) .(3.5) k k i i=2
Estimates i of the i are obtained by maximizing designweighted loglikelihood. The resulting predictions ik are used to form Yi GREG = U yik + yik = m sd k ( yik - yik), for i =1, 2, ..., I . Another development is the application of GREG reasoning to estimation for domains, as in Lehtonen, Srndal and Veijanen (2003, 2005) and Myrskyl (2007). Mixed models are used in the first two of these papers to assistthenonlinearGREG.Let Ua beadomain, U a U , whosetotal Yia =Ua yik wewishtoestimate, i =1, 2, ..., I . The 2005 paper derives the predictions for the nonlinear GREG from the logistic mixed model stating that for k a U Ex ( yik |x k uia ) = exp( x ia ) k
I exp(x ) 1 + k ia i= 2
(3.6)
(3.4)
where the model mean mk and the model variance Vx( yk | x ) aregivenappropriateformulations. k One application of (3.4) is when m k = m x , is a ( k ) specifiednonlinearfunctionin x . Havingestimated by k , the fitted values needed for Y GREG in (3.1) are ) for k . For example, if the modeler yk = m x , ( k U specifies log mk = a + bxk, thepredictionsforusein(3.1) are,followingparameterestimation, yk = exp(a + bxk). Other applications of (3.4) include generalized linear , modelssuchthat g (m k ) = xk foraspecifiedlinkfunction g ( and Vx ( yk | xk ) = v(mk) is given an appropriate ), , structure.Weestimate by thefittedvaluesneededfor the nonlinear GREG estimator (3.1) are ). yk = m k =g -1( x Forexample,usingalogisticassisting k model, x = logit(m k ) = log(m k /(1 - m )), and k k )). yk = m k = exp( x ) /(1 +exp( x k k Lehtonen and Veijanen (1998) examine the case of a categorical study variable with I classes, i =1, 2, ..., I , yik =1 if elementkbelongs to categoryi, and yik =0 if
with ia = i +uia , where uia is a vector of domain specificrandomdeviationsfromthefixedeffectsvector . i NonlinearGREGsassistedbymodelssuchas(3.5)and (3.6) require model fitting for everyyvariable separately there is no uniformly applicable weight system. However, the question arises: Are there examples of nonlinear GREGssuchthatthepracticaladvantagesoflinearGREG are preserved, that is, a linearly weighted form with calibratedweightsindependentthe yvariable.Theansweris intheaffirmative.Twodirectionsinrecentliteratureareof interestinthisregard: Breidt and Opsomer (2000), Montanari and Ranalli (2005) consider modelassisted local polynomial GREG estimators, for the case of a single continuous auxiliary variable with values x known for all k U . Several k choices havetobe madeinthe process:(1)the orderqof thelocalpolynomialexpression,(2)thespecificationofthe kernel function, and (3) the value of the band width. The resulting estimator can be expressed in terms of weights calibratedwithrespecttopopulationtotalsofthepowersof j x , sothat s wk xkj =U xk for j =0, 1, ..., q. k Breidt, Claeskens and Opsomer (2005) develop a penalizedsplineGREGestimatorfor asingle xvariablethe assisting model is m( x = b 0 + b1x + ...+ b qx q + ) q q q K j=1b q + j ( x - kj) + , where (t ) =t if t >0 and 0 + otherwise, q is the degree of the spline, and the kk are suitably spaced knots, for example, uniformly spaced
105 withweightsmadetoconfirmcomputableaggregates.This conceptual difference will sometimes lead to different estimatorsinthetwoapproaches. The calibration approach has considerable generality it can deal with a variety of conditions: complex sampling designs,adjustmentsfornonresponseandframeerrors.This section,however,focusesonthebasicconditionsinSection 2: single phase sampling and full response. The notation remains as in Section 2. The material available for estimatingthe population total Y =U yk is: (i) the study variable values y observed for k s, (ii) the known k design weights d k =1/k for k U , and (iii) the known vector values x for k U (or an imported total x ). U k k These simple conditions prevail in Deville and Srndal (1992) and Deville, Srndal and Sautory (1993), papers whichgavetheapproachanameandinspiredfurtherwork. Even though the background is simple, calibration raises severalissues,someofthemcomputational,asreviewedin Section5. The objective in Sections 4.2 and 4.3 is to determine weights w to satisfy the calibration equation k s wk xk =U x , then use them to form the calibration k estimatorof Y as YCAL = swk yk, whichwecanconfront with the unbiased HorvitzThompson estimator by writing YCAL = YHT + s( wk -dk ) yk. It follows that the bias of Y is E (YCAL) - Y = E ( s( wk -d k ) yk). Meeting the CAL objective of near design unbiasedness requires E ( s( wk - dk ) yk) 0, whatevertheyvariable.Evidently, thecalibrationshouldstriveforsmall deviations wk -d k. The objective calibration for consistency with known population auxiliary totals can be realized in many ways. We can construct many sets of weights calibrated to the known x . Thissectionexaminesthisproliferationfrom U k two perspectives noted in the literature: the minimum distance method and the instrumental vector method. Yet another construction of a variety of calibrated weights is proposedinDemnatiandRao(2004). 4.2 Theminimumdistancemethod In this method, the calibration sets out to modify the initialweights d k = 1/p intonewweights w , determined k k tobeclosetothe dk. Tothisend,considerthe distance function Gk( w, d), defined for every w >0, such that Gk ( w, d ) 0, Gk( d , d ) =0, differentiable with respect to w, strictly convex, with continuous derivative g k ( w, d ) = Gk( w, d ) / w such that g k( d , d ) =0. Usually the distance function is chosen such that g k ( w, d ) =g ( w /d ) / qk, wherethe q aresuitablychosen k positive scale factors, g ( is a function of a single ) argument, continuous, strictly increasing, with g (1) = 0, g (1) =1. Let F (u ) = g -1(u ) be the inverse function of g ( ). Minimizing the total distance
sample quantiles of the x values. After estimation of the k ) b parameters, they obtain the predictions yk =m ( xk needed for the general GREG formula (3.1). The authors pointoutthattheresultingGREGestimatoriscalibratedfor the parametric portion of the model, that is, j j s wk xk =U xk for j =0, 1, ..., q, and also for the truncatedpolynomialtermsinthemodelaslongatheyare leftunpenalized. We can summarize GREG estimation as follows. The linear GREG has practical advantages for large scale statistics production: It can be expressed as a linearly weighted sum of y values with weights calibrated to k x , the weights are independent of the y values and U k k may be applied to all yvariables in the survey. It is sufficient to know a population auxiliary total xk , U imported from a reliable source. Nonlinear GREG may givea considerablyreducedvariance,asaresultofthemore refined models that can be considered when there is completeauxiliaryinformation(known x forall k ) U k near design unbiasedness is preserved. Certain nonlinear GREGscanbewrittenaslinearlyweightedsums. In academic exercises with artificially created populations and relationships, one can provoke situations where a nonlinear GREG has a large variance advantage over a linear GREG. Such experiments are important for illustration.However,tomeetthedailyproductionneedsin nationalstatisticalagenciesfarfetchednonlinearGREGs seemtobeoffairlyremoteinterestatthispointintimethe assisting models for GREG must meet requirements of robustness and practicality. The attraction of a minor reductionofthesamplingvarianceissweptawaybyworries about other(nonsampling)errorsandtroublesinthe daily productionprocess. TheprogressionfromlineartononlinearGREGcreates opportunities and generates questions. What is the most appropriateformulationofthemodelexpectation mk ? How sensitivearetheresults tothespecification ofthe variance partoftheassistingmodel?Towhatextentiscomputational efficiency an issue? Further research will respond more fullytothesequestions.
4. Thecalibrationapproachtoestimation
4.1 Calibrationunderbasicconditions A crucial step in the GREG approach reviewed in the previoussectionistoproducepredictedvalues y through k the fit of an assisting model. By contrast, the calibration approach,asdefinedinSection1.1,doesnotreferexplicitly to any model. It emphasizes instead the information on which one can calibrate. A key element of calibration thinking is the linear weighting of the observedyvalues,
106
Gk ( wk , dk ) subject to the calibration equation s s wk xk =U x leads to wk =d k F ( qk x ), where is k k obtainedasthesolution(assumingoneexists)of
s d k xk F (qk xk ) = xk. U
(4.1)
Theweightshaveanoptimalityproperty,becauseaduly specifiedobjectivefunctionisminimized,butitisaweak optimality in the sense that there are many possible specificationsofthedistancefunctionandthescalefactors q . k Much attention has focused on the distance function Gk ( wk , d k ) = ( wk -d k ) 2 / 2d k qk. It gives g k ( wk , d k)= ( wk /d k - 1) / qk g ( w / d ) = w / d -1 F (u ) = g -1(u ) = 1 +u. Theterm the linear case is thus appropriate. The task is then to minimize the chisquare distance 2 s( wk -d k ) / 2dk qk, subject to s wk xk =U x . k Equation(4.1)reads s d k xk (1 + qk x ) =U x , whichis k k easilysolvedfor Theresultingestimatorof Y =U yk is . = swk yk with weights wk =d k g k given by (3.3). YCAL That is, YCAL =YGREG as given by (3.2), and the residuals that determine the asymptotic variance are Ek = yk - x B q as given in Section 3.2. Some negative k U weights w mayoccur. k ThelinearGREGestimatorimpliesweightsthathappen to be calibrated (to x ), and the opposite side of the U k samecoinsaysthatthelinearcaseforcalibration(withchi square distance) brings the linear GREG estimator. The tendency in some articles and applications to intertwine GREG thinking and calibration thinking stems from this fact. Many successful applications of the use of auxiliary information stem, in any case, from this linearity on both sidesofthecoin.TheCanadianLabourForceSurveyisan example, and an interesting recent development for that surveyisthe use of composite estimators, with part ofthe information coming from the survey results in previous months,asdescribedinFullerandRao(2001). Thecalibrationequationissatisfiedforanychoiceofthe positivescalefactors q in(4.1).Asimplechoiceis qk =1 k for all k. But it is not always the preferred choice. For example, if there is a single, always positive auxiliary variable, and x =xk, then many will intuitively expect k YCAL = swk yk to deliver the usual ratio estimator U xk ( s d k yk ) /( d k xk ), and it does, but by taking s qk =xk1, not qk =1. Another distance function of considerable interest is Gk ( wk , d k ) = {wk log( wk /d k ) - wk +d k }/ qk. It leads to F (u ) = g -1(u ) =exp(u ), theexponentialcase.Then(4.1) reads s d k xk exp(qk x ) = U x . Numeric methods are k k required to solve for , to obtain the weights ). wk =dk exp(qk x Nonegativeweights w willoccur. k k Deville and Srndal (1992) show that a variety of distance functions satisfying mild conditions will generate
asymptoticallyequivalentcalibrationestimators.Alternative distance functions are compared in Deville, Srndal and Sautory(1993),SinghandMohl(1996),Stukel,Hidiroglou andSrndal(1996).Somedistancefunctionswillguarantee weightsfallingwithinspecifiedbounds,soastoruleouttoo large or too small (negative) weights. Changes in the distance function will often have minor effect only on the varianceofthecalibrationestimator YCAL = swk yk, even if the sample size is rather small. Questions about the existence of a solution to the calibration equation are discussedinThberge(2000). 4.3 Theinstrumentvectormethod An alternative to distance minimization is the instrumental vector method, considered in Deville (1998), Estevao and Srndal (2000, 2006) and Kott (2006). It can alsogeneratemanyalternativesetsofweightscalibratedto thesameinformation. We can consider weights of the form wk =d k F ( zk), where zk is a vector with values defined for k s and sharingthedimensionofthespecifiedauxiliaryvector x , k and the vector is determined from the calibration equation s wk xk =U x . The function F ( plays the ) k same role as in the distance minimization method several choices F ( are of interest, for example, F (u ) = 1+u ) andF (u ) =exp(u ). Opting for the linear function F (u ) = 1 +u, we have w = d k (1 + zk ). Itisaneasyexercisetodetermine to k satisfy the calibration equation s wk xk =U x . The k resultingcalibrationestimatoris
YCAL = s wk yk wk = d k (1 + zk), =
xk - s dk x k
) ( d z x )
s k k k
-1
(4.2)
Whatever the choice of zk, the weights wk = d k (1 + zk ) satisfy the calibration equation. The standard choice is z k =x . In particular, setting z k =qk xk, for k specified q , givestheweights(3.3). k Even deliberately awkward choices for zk give surprisingly good results. For example, let x be a single k p continuous auxiliary variable, and zk =ck xk -1. Suppose p =3, and ck =1 for 4 elements only, chosen at random from n =100 elementsinarealizedsamples,and ck =0 for the remaining 96. The nearunbiasedness of YCAL = sd k (1 + zk )yk is still present.Even withsucha sparse zvector, the increase in variance, relative to better choicesof zk, maynotbeexcessive. When both sampling design and xvector are fixed, EstevaoandSrndal(2004)andKott(2004)notethatthere isanasymptotically optimal zvectorgivenby
z k = z 0k = d k-1s( d k d l - d kl ) l x l
107 MontanariandRanalli(2003,2005).Themotivatingfactor is that complete auxiliary information allows a more effectiveuseofthe x knownforevery k U thanwhatis k possible in modelfree calibration, where a known total x is sufficient. The weights are required to be U k consistent with the computable population total of the k predictions y , derived via an appropriate model formulation.Thustheweightsystemmaynotbeconsistent withthe known populationtotalofeachauxiliary variable, unless there is special provision to retain this property. Modelcalibrationstillsatisfiesallthreeparts,(a)to(c),of thedefinitionofcalibrationinSection1.1inparticular,the estimatorsarenearlydesignunbiased. Consider a nonlinear assisting model of the type (3.4). , We estimate the unknown parameter by leading to ) computed with the aid of fitted values yk = m k = m x , ( k the x knownforall k U . Itfollowsthatthepopulation k size N isknownandshouldbebroughttoplayasignificant role in the calibration. If minimum chisquare distance is used, we find the weights of the modelcalibration estimator YMCAL = s wk yk by minimizing ( wk -dk ) 2 /(2d k qk), for specified q , and d k = 1/ p , s k k subjecttothecalibrationequations
s wk = N s wk yk = yk. U
where d l is the inverse of the second order inclusion k probability pkl = P (k & l s), assumed strictly positive. The resulting calibration estimator, YCAL = sd k (1 + z k ) yk, is essentially the randomization 0 optimal estimator due originally to Montanari (1987) and discussedbymanysincethen. AnderssonandThorburn(2005)viewthequestionfrom the opposite direction and ask: In the minimum distance method, can a distance function be specified such that its minimization will deliver the randomizationoptimal estimator? They do find this distance not entirely surprisingly, it is related to (but not identical to) the chi squaredistance. 4.4 Doescalibration needanexplicitlystatedmodel? ThecalibrationapproachaspresentedinSections4.2and 4.3 proceeds by simply computing the weights that reproducethespecifiedauxiliarytotals.Thereisnoexplicit assisting model, unless one were to insist that picking certainvariablesforinclusioninthevector x amountstoa k serious modeling effort. Instead, the weights are justified primarilybytheirconsistencywiththestatedcontrols.Early contributionsreflectthisattitude,fromDeming(1943),and continuing with Alexander (1987), Zieschang (1990) and others.Thisbegsthequestion:Isitneverthelessimportantto motivate such modelfree calibration with an explicit model statement? It is true that statisticians are trained to think in terms of models, and they feel more or less compelled to always have a statistical procedure accompanied by a model statement. It may indeed have some pedagogical merit, also in explaining calibration, to state the associated relationship of y to x, even if it is as simpleasastandardlinearmodel. But will a stated model help the users and practitioners betterunderstandthecalibrationapproach?Tomostofthem the approach is perfectly clear and transparent anyway. They need no other justification than the consistency with stated controls. Will a search for the true model with the true variance structure bring significantly better accuracy for the bulk of the many estimates produced in a large governmentsurvey?Itisunlikely. The next section deals with modelcalibration. For that variety, proposed by Wu and Sitter (2001), modeling has indeedanexplicitandprominentrole.Theseauthorscallthe linear calibration estimator, YCAL = s wk yk with weights w given by (3.3), a routine application without k modeling. The descriptionisappropriateinthatallthatis necessary is to identify the xvariables with their known populationtotals. 4.5 Modelcalibration The idea of modelcalibration is proposed in Wu and Sitter (2001) and pursued further in Wu (2003) and
(4.3)
Forsimplicity,letustake qk =1 forall kwederivethe calibratedweights,rearrangetermsandfindthatthemodel calibrationestimatorcanbewrittenas

YMCAL = N{ ys d + ( yU - ys d )Bs d} %
(4.4)
where ysd = s d k yk / s d k ysd = s d k yk / sd k, and % Bs d =
d k ( yk - ys d ) yk
s k
d ( yk - ys d )2.
%s The regression implied by Bd is one of observed y values on predicted yvalues. The idea of this regression would hardly occur to the modeler is his/her attempts to structure the relation between y and x , but it proves k k effectiveinbuildingthecalibrationestimator.WuandSitter (2001)presentevidencethat (YMCAL - Y ) / N =
% % d k Ek - U Ek
N + O p( n -1)
% % % with Ek = yk - yU - ( mk - mU ) BU, where BU = 2 (U (m k - mU ) yk ) / U (mk - m ) , and mU = U m / N . U k % The coefficient B may not be near one even in large U samples. It expresses a regression of y on its assisting k model mean m k = m xk , That is, Y ( ). MCAL canbe viewed asaregressionestimatorthatusesthemodelexpectation m k % as the auxiliary variable, leaving E as the residuals that k determinetheasymptoticvarianceof YMCAL.
108
Howdoesthisasymptoticvariancecomparewiththatof the nonlinear GREG construction (3.1) for the same non linearassistingmodelandthesame yk = mk? Formula(3.1) impliesaslopeequaltounityintheregressionbetween y k and yk = mk viewed in that light, Y GREG is a difference estimatorratherthanaregression estimatorand henceless sensitive to the pattern in the data. The nonlinear GREG Y GREG isingenerallessefficient than Y MCAL. (Itisofcourse possible to modify Y to also account for the GREG informationcontainedintheknownpopulationsize N.) Ontheotherhand,comparedwiththelinear(modelfree) calibration estimator YCAL = s wk yk with weights as in (3.3),themodelcalibrationestimator Y MCAL givenby(4.4) may have a considerable variance advantage but implies a loss of the practical advantages of a consistency with the known population total x and a multipurpose weight U k systemapplicabletoall yvariables.The yvaluesin(4.4)are linearlyweighted,buttheweightsnowalsodependonthe y values. It is thus debatable if Y MCAL is a bona fide calibrationestimator. In an empirical study, Wu and Sitter (2001) compare YMCAL = s wk yk, calibrated according to (4.3), with the nonlinearGREG, Y GREG =U y k + sd k ( y k - yk ) givenby (3.1), for the same nonlinear assisting model and same k yk = m . The study confirms that Y MCAL has a variance advantageoverthenonlinear Y . Theycreatedafinite GREG population U of size N =2,000 with values ( yk , x ), k =1, ..., 2,000, suchthat log( yk )= 1 + xk + e k k the 2,000 values x are realizations of the Gamma(1,1) k randomvariable,and ek isanormallydistributederror.The auxiliaryinformationconsistsofthepopulationsizeNand theknownvalues x for k =1,..., 2,000. Repeatedsimple k random samples of sizen= 100 were taken the assisting model for both estimators was the loglinear Ex ( yk | xk )= m with log(m k ) = a + bxk. This model was k fit for each sample, using pseudomaximum quasi likelihoodestimation.Thefittedvalues yk = exp(a + bxk) wereusedtoformboth Y and Y . Thesimulation MCAL GREG variancewasmarkedlylowerfor Y MCAL. (ThelinearGREG (3.2),identicaltothe modelfree calibration estimator,was alsoincludedintheWuandSitterstudynotsurprisingly,it is evenless efficientthanthe nonlinearGREG, underthe strongly nonlinear relationship imposed in their experiment.) Montanari and Ranalli (2005) provide further evidence, for several artificially created populations, on the . comparison between Y MCAL and the nonlinear Y GREG Their assisting model, yk = m k + e , is fitted via k nonparametric regression (local polynomial smoothing), yielding predictions yk = m for k U . With this type of k modelfit,thepredictions yk = m arehighlyaccurate.Not k surprisingly, the modelcalibration estimator Y MCAL
achieves only marginal improvement over the nonlinear . Y GREG Wecansummarizethe calibrationapproachasfollows: Theestimatorof Y =U yk hasthelinearlyweightedform Y = s wk yk. In linear (modelfree) calibration, the calibration equation reads swk x = x a known U k k population auxiliary total x is required, but complete U k auxiliary information (known x for all k ) is not U k requiredthesameweightscanbeappliedtoallyvariables (multipurpose weighting) the estimator is identical to the linearGREGestimator(butderivedbydifferentreasoning). In modelcalibration, theassisting model mean mk isnon linear in x complete auxiliary information is usually k required the calibration constraints include the equation s wk yk =U yk the weights w depend on the y k k values,implyingalossofthemultipurposeproperty.
5. Computationalaspects,extremeweights andoutliers
The computation of calibrated weights raises important practical issues, discussed in a number of papers. All computation must proceed smoothly and routinely in the large scale statistics production of a national statistical agency.Undesirable(orundulyvariable)weightsshouldbe avoided. Many practitioners support the reasonable requirement that all weights be positive (even greater than unity)andthatverylargeweightsshouldbeavoided. A few of the weights computed according to (3.2) can turn out to be quite large or negative. Huang and Fuller (1978) and Park and Fuller (2005) proposed methods to avoidundesirableweights. In the distance minimization method, the distance function can be formulated so that negative weights are excluded, while still satisfying the given calibration equations. The software CALMAR (Deville, Srndal and Sautory1993)allowsseveraldistancefunctionsofthiskind. An expended version, CALMAR2, is described in LeGuennec and Sautory (2002). Other statistical agencies havedevelopedtheirownsoftwareforweightcomputation. Among those are GES (Statistics Canada), CLAN97 (Statistics Sweden), Bascula 4.0 (Central Bureau of Statistics, The Netherlands), gCALIBS (Statistics Belgium). These strive, in different ways, to resolve the computationalissuesarising.The userneedsto consultthe users guideineach particularcaseto seeexactly howthe computationalissues,includinganavoidanceofundesirable weights,arehandled. GES uses mathematical programming to minimize the chisquaredistance,subjecttothecalibrationconstraintsas wellastoindividualboundsontheweights,sothattheywill satisfy Ak wk Bk for specified Ak , B . Bascula 4.0 is k
109 distribution function must first be estimated. Before calibration became popular, several papers considered the estimationofquantiles,withorwithouttheuseofauxiliary information. More recent articles have turned to the calibration approach for the same purpose, including Kovaevi(1997),WuandSitter(2001),Ren(2002),Till (2002), Harms (2003), Harms and Duchesne (2006) and Ruedaetal.(2007).Asthesepapersillustrate,thereismore than one way to implement the calibration approach. The nonsmooth character of the finite population distribution function causescertain complexitiestheseareresolvedby differentauthorsindifferentways. Let D( denote the Heaviside function, defined for all ) realzsothat D( z ) =1 if z 0 and D( z ) =0 if z <0. The unknowndistributionfunctionofthestudyvariable y is Fy (t ) = 1 D(t - yk). N U (6.1)
described in Nieuwenbroek and Boonstra (2002). The software gCALIBS, described in Vanderhoeft, Waeytens and Museux (2001), Vanderhoeft (2001), uses generalized inverse (the MoorePenrose) for the weight computation consequently one need not be concerned about a possible redundancyintheauxiliaryinformation. InBankier,Houleand Luc(1997)the objectiveistwo fold:tokeepthecomputedweightswithindesirablebounds, and to drop some xvariables to remove nearlinear dependencies. Isaki, Tsay and Fuller (2004) consider quadratic programming to obtain both household weights andpersonweightsthatliewithinspecifiedbounds. An intervention with the weights (so as to get rid of undesirable weight values)raises thequestion how farone can deviate from the design weights d without k compromising the desirable feature of nearly design unbiasedestimation.Anideathathasbeentriedistomodify thesetofconstraintssothattolerancesarerespectedforthe differencebetweentheestimatorfortheauxiliaryvariables and the corresponding known population totals. Hence, Chambers(1996)minimizesacostridgedlossfunction. Outlyingvaluesintheauxiliaryvariablesmaybeacause ofextremeweights.Calibrationinthepresenceofoutliersis discussed in Duchesne (1999). His technique of robust calibrationmayintroduceacertainbiasintheestimatesit may, however, be more than offset by a reduction in variance. When the set of constraints is extended to make the weights restricted to specified intervals, a solution to the optimizationproblemisnotguaranteed.Theexistenceofa solution is considered in Thberge (2000), who also proposesmethodsfordealingwithoutliers.
The aquantile of the finite population is defined as Qya = inf{t | Fy(t ) a The auxiliary variable xj, taking }. values xjk, has the distribution function Fxj (t )= (1/ N )U D(t -x jk) with aquantile denoted Qxja, ) j =1, 2, ..., J . A natural estimator of Fy(t based on the designweights d k = 1/p is k Fy (t ) = 1 d k D(t - yk). sd k s
) Acalibrationestimator Fy(t oftakestheform FyCAL(t ) = 1 wk D(t - yk) swk s (6.2)
6. Calibrationestimationformore complexparameters
Thecalibrationapproachadaptsitselftothe estimationof morecomplexparametersthanapopulationtotal.Examples arereviewedinthissection.Singlephasesamplingandfull responsecontinuetobeassumedthenotationremainsasin Section 2. One example is the estimation of population quantiles (Section 6.1), another is the estimation of functions of totals (Section 6.2). Other examples in this category, not reviewed here, are Thberge (1999), for the estimation of bilinear parameters, and Tracy, Singh and Arnab (2003), for calibration with respect to second order moments. 6.1 Calibrationforestimationofquantiles The median and other quantiles of the finite population areimportant descriptive measures,especiallyineconomic surveys. To estimate quantiles, the finite population
wheretheweights w aresuitablycalibratedtoaspecified k auxiliary information then from FyCAL(t) we obtain the aquantile estimator as Qya = inf{t | FyCAL(t ) a A }. CAL(t). formulaanalogousto(6.2)holdsfor Fxj Without explicit reference to any model, Harms and Duchesne (2006) specify the information available for calibration as a known population size, N, and known population quantiles Qxja for j =1, 2, ..., J . The complete auxiliary information, with values x =( xk1, ..., xkJ ) k known for k U , is not required. (But in practice, the completeinformation would usuallybe necessary,because accuratequantilesofseveralxvariablesarenotlikelytobe importablefromoutsidesources.)Theydeterminethe w to k minimizethechisquaredistance s( wk -d k ) 2 / 2d k qk, for specified q , subjecttothecalibrationequations k swk = N Qx CAL,a = Qx a , j = 1, 2, ...,J
j j
forsuitablydefinedestimates QxjCAL,a. Now,ifwewereto }, specify Qx j CAL, a = inf{t | FxjCAL(t ) a thenitisingeneral not possible to find an exact solution of the calibration
110
problemasstated.Instead,HarmsandDuchesnesubstitute smoothed estimators, called interpolated distribution estimators, of the distribution functions Fxj (t ), j =1, 2, ..., J . They replace D( by a slightly ) modified function. Weights w can now be obtained, as k well as a corresponding estimated distribution function -1 FyCAL(t) finally, Qya isestimatedas Qya = FyCAL(a ). Theresultingcalibratedweights w allowustoretrieve k the known population quantiles of the auxiliary variables. This is reassuring one would expect such weights to producereasonableestimatorsforthequantilesofthestudy variable y.Moreover,inthecaseofasinglescalarauxiliary variablex,theresultingcalibrationestimatordeliversexact populationquantiles forywhen therelationshipbetweeny and x isexactlylinear,thatis,when yk = bxk forall k U . An idea involving smoothed distribution functions is also usedinTill(2002). The computationally simpler method of Rueda et al. (2007) is an application of modelcalibration, in that they calibrate withrespect toapopulationtotalofpredicted y values. Complete auxiliary information is required. Using the known x , compute first the linear predictions k -1 yk = x k for k , with =( sd k qk x k x ) U k ( sd k qk x yk), where d k = 1/p andthe q arespecified k k k scalefactors. The weights w are obtainedby minimizing k the chisquare distance subject to calibration equations statedintermsofthepredictions,soastohaveconsistency at J arbitrarilychosenpoints t j, j =1, ..., J : 1 wk D(t j - yk ) = Fy(t j), j = 1, ...,J N s where Fy (tj) is the finite population distribution function ofthepredictions y , evaluatedat tj. Itissuggestedthata k fairly small number of arbitrarily selected points tj may suffice,saylessthan10.Oncethe w aredetermined,the k aquantile estimate is obtained from Fy CAL(t ) = (1/ N ) swk D(t - yk). Quantile estimation providesa good illustrationthat the calibration approach can be carried out in more than one way when somewhat more complex parameters are being estimated. Both methods mentioned give nearly design unbiased estimation. The Harms and Duchesne (2006) weightsaremultipurpose,independentofthe yvariableby contrast,themethodofRuedaetal.(2007)requiresanew setofweightsforeverynew yvariable.Empiricalevidence, by simulation, suggests that both methods compare favourablywiththeearlierquantileestimationmethods,not based explicitly on calibration thinking (but on the same auxiliaryinformation). An extension of the calibration approach to the estimation of other complex parameters, such as the Gini coefficient,issketchedinHarmsandDuchesne(2006).
6.2 Calibrationforothercomplexparameters Plikusas (2006), and Krapavickait and Plikusas (2005) examine calibration estimation of certain functions of population totals. (Their term nonlinear calibration signifiesnonlinearfunctionoftotalsIdonotuseithere.) Asimpleexampleistheestimationofaratiooftwototals, R =U y1k / U y2k, where y and y arethevaluesfor 1k 2k element k of the variables y and y , respectively. (The 1 2 distributionfunction(6.1)isineffectalsoofratiotype,with y2k =1, and N =U1 as the denominator total.) These authors examine the calibration estimator RCAL = s wk y1k / swk y2k. Its weights w , common to k the numerator and the denominator, are determined by calibrationtoauxiliaryinformationstatedasfollows:There isoneauxiliaryvariable, xk, for yk, andanother, x k, for 1 1 2 y k the ratio of totals R0 =U x1k /U x2k is a known 2 value,byacompleteenumerationatapreviousoccasionor fromsomeotheraccuratesource.Theproposedcalibration equation is swk ek =0, where ek = x1k -R0 x2k. Because U ek =0, the weights, by minimum chisquare distance, are
wk = d k 1 -
( d e ) ( d e )
s k k 2 s k k
-1
ek .
These weights correctly retrieve the known ratio value R setting y1k =x1k and y2 k = x2k in R , wehave 0 CAL
s wk x1k s wk x2k
- R = 0
swk ek = 0. swk x2k
The empirical evidence in Plikusas (2006), and Krapavickait and Plikusas (2005) suggests that their calibration estimator compares favourably(lower variance, while maintaining near design unbiasedness) with other estimators, derived through other arguments than cali bration,whilerelyingonthesameauxiliaryinformation.
7. Calibrationcontrastedwithotherapproaches
As manyhavenoted,usersviewcalibrationasasimple andconvincingwaytoincorporatingauxiliaryinformation, for simple parameters (Section 4), as for more complex parameterssuchasquantiles,ratiosandothers(Section6). Simplicity and practicality are undeniable advantages, but asidefromthat,iscalibrationalsotheoreticallysuperior? Arethereinstanceswherecalibrationcanbeshowntogive more accurate and/or more satisfactory answers on questionsofimportance,whencontrastedwithotherdesign basedapproaches? Section4.5gaveoneindicationthatcalibrationthinking mayhaveanadvantageoverGREGthinking,inthatmodel calibration may give more precise estimates than the
111
in an attempt to borrow strength for Y GREG by letting it a dependalsoon ydatafromoutsidethedomain.Bycontrast, Y CAL reliesexclusivelyon ydatainthedomain,andthisis q in effect better. Estevao and Srndal (2004) show that Y CAL with z k =z0Ck has smaller (asymptotic) variance a than Y GREG, nomatterhow s ischosen.Bringingin ydata % a from the outside does not help calibration thinking and regressionthinkingdonotagree.
nonlinear GREG, for the same assisting model. The following Section 7.1 gives another example where calibration reasoning and GREG reasoning give diverging answers,withanadvantageforthecalibrationmethod. 7.1 Anexampleindomainestimation The example in this section, from Estevao and Srndal (2004), shows, for a simple practical situation, a conflict between the results of GREG thinking and calibration thinking. The context is the estimation of theytotal for a subpopulation(adomain). A probability sample s is drawn from U = 1, 2,..., k , ..., N} the known design weights are { d k =1/ k. Let Ua be a domain U a U . The domain indicatoris dak withvalue d ak =1 if k U a and d ak =0 if not. The target of estimation is the domain total Ya =U yak, where yak = dak yk, and y is observed for k k s. The HorvitzThompson estimator YHT = sd k yak, although design unbiased, has low precision, especially if the domain is small the use of auxiliary information will bring improvement. An auxiliary vector value x is k specifiedforevery k U . As is frequently the case in practice, the elements belonging to a domain of interest are not identified in the sampling frame. (If they are, some very powerful information is available from the start, but frequently real world conditions are not that favourable.) But suppose elements in a larger group UC are identifiable U a U C U . For example, supposey is income and UC aprofessionalgroupspecifiedforthepersonslistedin the frame, while Ua is a professional subgroup not identifiedintheframe.Wecanidentifythesamplesubsets sC = s U C and sa = s U a, and we can benefit from knowing the total x , estimable without bias by U Ck d k x , where xCk = d x , and d istheinformation s Ck k Ck Ck groupindicator: dCk =1 if k U C and dCk =0 ifnot.The domainauxiliarytotal x isunavailable,because Ua is U ak not identified. Calibration to satisfy s wk xCk =U x Ck gives the nearly design unbiased estimator YaCAL =U wk yak, where wk = d k (1 + zk ), with = ( U x Ck - sd k x ) ( sd k z k x ) -1. The Ck Ck asymptoticallyoptimalinstrumentforthegivenvector x is k (seeSection4.3) z k = z0Ck = d k-1 ls( d k d l -d k l ) x l. C By contrast, regression thinking for the same auxiliary information leads to YaGREG = s d k yak + (U x - Ck B sd k xCk ) s%d, also nearly design unbiased, where the regressioncoefficient B s%d =( s% d k x k x )-1 s d k x yk isthe % k k resultofaweightedleastsquaresfitat asuitablelevel,using % % all (when s = s ) or part (when s s ) of the data points ( y , x ) availablefor k s. k k For example, the modeller may opt for a regression fit % extendingbeyondthedomain(sothat s sa = s U a),
8. Calibrationestimationinthepresence ofcompositeinformation
As the preceding sections have shown, many papers choosetostudyestimationfordirect,singlephasesampling of elements, without any nonresponse. The information available for calibration is simple thek:th element of the finite population U ={1, 2, ..., k , ...,N } has an associated auxiliaryvectorvalue x . k However, in an important category of situations, the auxiliary information has composite structure. The complexity of the information increases with that of the samplingdesign.Indesignswithtwoormorephases,orin two or more stages, the information is typically composed of morethan onecomponent,reflectingthefeatures ofthe design.Theinformationisstatedintermsofmorethanone auxiliaryvector.Forexample,intwostagesampling,some informationmaybeavailableaboutthefirststagesampling units(theclusters),otherinformationaboutthesecondstage units(theelements). Consequently, estimation by calibration (or by any alternativemethod)musttakethecompositestructureofthe information systematically into account. The total infor mation has several pieces the calibration can be done in morethanoneway.Allrelevantpiecesshouldbetakeninto account, for best possible accuracy in the estimates. To accomplishthisinageneraloroptimalwayisnotatrivial task.Calibrationreasoningoffersoneway. Regression reasoning, with a duly formulated assisting model,isanalternativeway,butitwillstrikesomeusersas more roundabout. Hence, surveys that allow composite auxiliary information bring further perspectives on the contrastbetweencalibrationthinkingandGREGthinking. Twophase sampling and twostage sampling are discussed in this section. Another example of composite information occurs for nonresponse bias adjustment, as discussedinSection9. Another aspect of composite information occurs when the objective is to combine information from several surveys.This,too,canbeawaytoaddstrengthandimprove accuracy of the estimates. It is a motivating factor (in additiontotheuserorientedmotivetoachieveconsistency
112
among surveys) in the previously mentioned repeated weighting methodology of the Dutch statistical agency. Combined auxiliary information for GREG estimation is consideredinMerkouris(2004). 8.1 Compositeinformationfortwophasesampling designs Double sampling refers to designs involving two probability samples, s and s , from the same population 1 2 U ={1, ..., k , ..., N }. Auxiliary data may be recorded for bothUand s, the study variable values y are recorded 1 k only for k s2 with an objective to estimate Y =U yk. Hidiroglou (2001) distinguishes several kinds of double sampling: In the nested case (traditional two phase sampling),thefirstphasesample s isdrawnfrom U, the 1 second phase sample s is a subsample from s, so that 2 1 U s1 s2. Twononnested cases can be distinguished: In the first of these, s is drawn from the frame U s 1 1 2 from the frame U2, where U and U2 cover the same 1 population U the sampling units may be defined differently for the two frames. In the second nonnested case, s and s aredrawnindependentlyfrom U. 1 2 Toillustratehowcompositeinformationintervenesinthe estimation,considerthenestedcase.Thedesignweightsare p d1k = 1/pk (s sampledfromU) d 2 k = 1/ p 2 k ( 2k = p s 1 1 k 1 insubsampling s from s ). Thecombineddesignweight 2 1 is d k =d1k d 2k. Thebasicunbiasedestimator Y = s2 d k yk canbeimprovedbyauseofauxiliaryinformation,specified hereattwolevels: Populationlevel:Thevectorvalue x isknown(givenin 1k the frame) for every k U , thus known for every k s1 andforevery k s2 U xk isaknownpopulation vector 1 total First sample level: The vector value x is known 2k (observed)forevery k s1, andtherebyknownforevery k s2 the unknown total x k is estimated without U 2 biasby s d1k x k. 2 1 How do we best take this composite information into account? In an adaptation of GREG thinking, Srndal and Swensson (1987) formulated two linear assisting models, thefirstonestatedintermsofthex vector,theotherone 1k alsobringsinthe x vector.Thetwomodelsarefittedthe 2k resulting predictions, of two kinds, are used to create an appropriateGREGestimator Y GREG of Y =U y k. Dupont(1995)makestheimportantpointthatthegiven composite information invites two different natural approaches: Besides the GREG approach, there is a calibration approachthatwilldeliverfinalweights w fora k calibration estimator YCAL = s2 wk yk. It is of interest to compare the results of the two approaches. Both of them allow more than one option: IntheGREGapproach,there are alterative ways of formulating the linear assisting
models with their respective variance structures. In the calibration approach, alternative formulations of the calibrationequationsarepossible. Forexample,atwostepcalibrationoptionisasfollows: First find intermediate weights w to satisfy 1k w1k x1k =U xk thenusetheweights w inthesecond s1 1 1k steptocomputethefinalweights w tosatisfy k
U xk 1 wk x k = s wk x = 1 k 2 1 w1k x k 2 s 1
where xk isthecombinedauxiliaryvector x x = 1k . k 2 x k Alternatively, in asingle step option, we determine the w directlytosatisfy k

U xk 1 w x = s2 k k d x . s 1 1k 2k
Thefinal weights w arein generalnotidenticalinthe k twooptions. Supposethat xk isanimported x total. U 1 1 Atcloserlook,thetwostepoptionrequiresmoreextensive information, because individually known values x are 1k requiredfor k s1, whereasitissufficientinthesinglestep option that they be available for k s2. Some variance advantage may thusbe expected from the twostep option, since s w1k x k isoftenmoreaccurate(asanestimatorof 2 1 x k ) than 1 d1k x k in the single step procedure. U 2 s 2 Nevertheless,thisanticipationisnotalwaysconfirmedthe single step method can be better, as when x and x are 1 2 weaklycorrelated. Dupont (1995) and Hidiroglou and Srndal (1998) examinelinksthat exist, notsurprisingly,betweenthetwo approaches. A GREG estimator, derived from assisting modelswithspecificvariancestructures,maybeidenticalto calibration estimator, if the weights of the latter are calibratedinacertainway.Inothercases,differencesmay besmall. The efficiency of different options depends in rather subtlewaysonthepatternofcorrelationamong y , xk and k 1 x k. For example, to what extent do x and x 2 1 2 complementeach other,to whatextentaretheysubstitutes for one another? In the GREG approach, it is difficult or even futile to pinpoint a variance structure that truly captures a reality behind the data. The calibration approach is more direct. Some of its possibilities are exploredinEstevaoandSrndal(2002,2006). 8.2 Compositeinformationintwostagesampling designs The traditional twostage sampling setup (clusters sampledatstageone,elementssubsampledwithinselected
113 Once the w are determined, the element weights Ii wk =d k |i wIi follow. Another reasonable integrated weighting is to impose si wk = N i wIi. For example, for single stage cluster samplingitimpliesthat theclusterweight w istheaverage Ii oftheelementweights w inthatcluster. k TwostagesamplingisalsothetopicinKim,Breidtand Opsomer (2005). They assume auxiliary information for clusters, viaasinglequantitativecluster variable xc )i, but ( noneforelements.TheydevelopandexamineaGREGtype estimator of the element total Y =U yk, i Y = iU I mi + isI d i (t - m ), where t isdesignunbiased i i i fortheclustertotal ti =Ui yk, and m isobtainedbylocal polynomial regression fit. The estimator can be expressed onthelinearlyweightedform,withweightsthatturnoutto becalibratedtothepopulationtotalsofpowersofthecluster variable xc )i. ( 8.3 Householdweightingandpersonweighting Some important social surveys set the objective to produce both household estimates and person estimates some study variables are household (cluster) variables, others are person (element) variables. Consequently, a number of papers have addressed the situation withsingle stage cluster sampling ( d k |i =1) and the integrated weighting that gives all members of a selected household equal weight, a weight alsoused for producing household statistics. A general solution for this weighting problem, when both household information and person information are specified, is to obtain the household weights w Ii calibrated as in equation (8.1) with d k |i =1, then take wk = wIi. Several articles focus on auxiliary vector values x k attributed topersons. Alexander(1987)derives weightsby minimizing chisquare distance, whereas Lematre and Dufour (1987) and Niewenbrook (1993) derive the integratedweightsviaaGREGestimator.TheLematreand Dufourtechniqueproceedsbyanindirectconstructionofan equal shares auxiliary vector value for all persons in a selected householdtheirresultis derivablefromthedirect procedureinSection8.2. The householdweighting/personweighting question is revisited in more recent papers. Some authors display calibrationthinking,othersGREGthinking.Isaki,Tsayand Fuller (2004) formulate the problem as one of calibrated weighting their weights respect both household controls and person controls no explicit assisting models are formulated.Bycontrast,SteelandClark(2007)proceedby theGREGapproach,withlinearassistingmodelstatements andaccompanyingvariancestructures.
clusters in stage two) has in common with twophase samplingthatthetotalinformationmayhavemorethanone component. There may exist (a) information at the cluster level (about the clusters) (b) information at the element levelforallclusters(c)informationattheelementlevelfor the selected clusters only. Here again, authors are of two different orientations: some exploit the information via calibrationthinking, othersfollowtheGREGthinkingroute. Estevao and Srndal (2006) develop calibration estimation for the traditional twostage setup, with composite information specified as follows: (i) for the cluster population UI, there is a known total I xc )i , U ( where xc )i isavectorvalueassociatedwiththecluster Ui, ( for i U I (ii) for the population of elements U = UUI U i, there is a known total xk , where the vector U i value x is associated with the element k U . Suppose k both cluster statistics and element statistics are to be produced in the survey: Both the cluster population total YI =UI y( c )i and the element population total Y =U yk aretobeestimated. Ifnorelationisimposedbetweenclusterweights w and Ii element weights w , the former are calibrated to satisfy k the latter to satisfy sI wIi x( c )i =UI xc)i, ( s wk x k =U x . (Here, s isthesample of clusters from k I U Is is the sample of elements from the cluster Ui i s = UsI s is the entire sample of elements.) Then i i YICAL = sI wIi y(c )i estimates the cluster population total Y, and YCAL = swk yk estimates the element population I total Y. Integrated weighting is often used in practice: A convenient relationship is imposed between the cluster weight w andtheweights w fortheelementswithinthe Ii k selected cluster. Two forms of integrated weighting are discussedinEstevaoandSrndal(2006). Oneof theseistoimpose wk =d k |i wIi, where dk |i isthe inverse of the probability of selecting element k within cluster i. (For example, in single stage cluster sampling, whenallelementskinasampledclusterareselected,then d k |i =1. Consequently wk =wIi is imposed, and all elements in the cluster receive the same weight for computing element statistics, and that same weight is also used for computing cluster statistics.) The calibration equation then reads s wk x k =U x k sI wIi si d k |i xk =U x . Theclusterweights w arenow k Ii derived by minimizing sI ( wIi -d Ii ) 2 /d Ii subject to the calibrationequationthattakesbothkindsofinformationinto account: s wi xc )i U xc )i I ( ( I I = . x s wIi s d k |i x k U k I i (8.1)
114
9. Calibrationfornonresponseadjustment
9.1 Traditionaladjustmentfornonresponse The context of many good theory articles is the simple one of Section 2, which includes total absence of nonresponse.Itisgood theoryforconditionsthatseldomor neveroccur.(Asanauthorofpapersinthatstream,Iamnot without guilt.) Practically all surveys encounter non response although undesirable, it is a natural feature, and theory should incorporate it, from the outset, via a perspectiveofselectionintwophases. In many surveys, nonresponse rates are extremely high today,comparedwithwhattheywere40yearsago,thatis, solowthatonecouldessentiallyignoretheproblem.Today, surveysamplingtheoryneedsmoreandmoretoaddressthe damaging consequences of nonresponse. In particular, one pressingobjectiveistoexaminethebiasandtotrytoreduce itasfaraspossible. A probability sample s is drawn from U = {1, 2, ..., k , ..., N} theknowndesignweightofelement k is d k =1/ k. Nonresponseoccurs,leavingaresponsesetr, a subset ofs the study variable value y is observed for k k r only.Theunknownresponseprobability ofelement k is Pr(k r|s )= k. Theunbiasedestimator Y = rd k fk yk isruledoutbecause f k =1/k isunknown.Tokeeptheidea of a linearly weighted sum, how do we then construct the weights?Unitnonresponseadjustmentbyweighting,based onnonresponsemodeling,hasalonghistory.Calibration offersanewerperspective. In what we may call the traditional procedure, the probability design weights d k =1/k are first adjusted for nonresponse and possibly for other imperfections such as outliers. The information used for this step is often a grouping of the sampled elements. Finally, if reliable populationtotalsareaccessible,theadjusteddesignweights aresubjectedtoacalibrationwithrespecttothosetotals. ThemethodologyoftheLabourForceSurveyofCanada, described in Statistics Canada (1998), exemplifies this widespread practice. A (modified) design weight is first computed for a given household, as the product of three factors.Theproductofthedesignweightandanonresponse adjustmentfactoriscalledthesubweight.Thesubweights aresubjectedinthefinalsteptoacalibrationwithrespectto postcensal, highly accurate estimates of population by age group, sex and subprovincial regions. The final weights meet the desirable objective of consistency, in regions within a province, with the postcensal estimates. The nonresponse bias remaining in the resulting estimates is unknownbutbelievedtobemodest. The traditional procedure is embodied in the estimator type Y = rd k (1/ k ) yk, where k hasbeenestimatedby in a preliminary step, using response (propensity) k
modeling.Whattheorydemandsofthestatisticianisnotan easytask, namely,to formulatethetrueresponse model, capable of providing accurate, nonbiasing values k. But areappliedinmanysurveysinanuncritical thefactors 1/k andmechanicalfashion,forexample,bystraightexpansion withinthestrataalreadyusedforsampleselection. The traditional procedure is apparent for example in Ekholm and Laaksonen (1991) and in Rizzo, Kalton and Brick(1996). Practitioners often act as if the resulting Y = )y (followingamoreorlessprobingresponse d k (1/ k k r modeling trying to get the k) is essentially unbiased, somethingwhichitisnot(unlesstheidealmodelhappensto bespecified)oneacts(forpurposesofvarianceestimation, forexample)asif p k isthetrueselectionprobabilityof k element k inasinglestepofselection,somethingwhichitis definitely not. This practice, with roots in the idyllic past, becomes more and more vulnerable as nonresponse rates continuetheirsurreptitiousclimb. Anunavoidablebiasresultsfromthereplacementof k by k. Decadesago,whenthetypicalnonresponsewasbut afewpercent,itwasdefendabletoignorethisbias,but with todays galloping nonresponse rates, the practice becomes untenable. By first principles, unbiased estimation is the goal, not an estimation where the squared bias is a dominating(andunknown)contributortotheMeanSquared Error.Wemustresolvetolimitthebiasasmuchaspossible. Calibrationreasoningcan helpinconstructinganauxiliary vectorthatmeetsthisobjective. 9.2 Calibrationfornonresponsebiasadjustment More or less contrasting with the traditional procedure are a number of recent papers that emphasize calibration reasoning to achieve the nonresponse adjustment. Recent referencesareDeville(1998,2002),Ardilly(2006),chapter 3,Skinner(1998),FolsomandSingh(2000),Fuller(2002), Lundstrm and Srndal (1999), Srndal and Lundstrm (2005)andKott(2006). Calibration reasoning starts by assessing the total available auxiliary information: information at the sample level (auxiliary variable values observed for respondents andfornonrespondents),informationatthepopulationlevel (known population auxiliary totals). The objective is to makethebestofthetwosourcescombined,soastoreduce bothbiasandvariance.Thedesignweightsaremodified,in one or two calibration steps, to make them reflect (i) the outcome of the response phase, (ii) the individual characteristics of the respondents, and (iii) the specified auxiliary information. Theinformation canbesummarized asfollows:
* Populationlevel:The vector value x isknown(specified k intheframe)forevery k U , thusknownforevery k s * andforevery k rU x isaknownpopulationtotal. k
115 probably more important than the choice of the weighting methodology. Let us examine the bias, when z k =x . We need to k comparealternative xk vectorsinordertofinallysettleone likely to yield the smallest bias. (I assume xk to be such that xk =1 forall k andsomeconstantvector , asisthe caseformany xk vectors,includingtheexamples1to5at the beginning of Section 2.) A close approximation to the bias of Y CAL is obtained by Taylor linearization as nearbias(YCAL ) = (U x k )(BU -B ), which involves a U difference between the weighted regression coefficient -1 BU =( U k x k x ) U x yk andthe unweighted one, k k k BU =(U x k x ) -1( U x yk). Unless all are equal, the k k k biascausedbythe differenceinthetworegression vectors maybesubstantial,eventhough xk isaseeminglygood auxiliary vector. This expression for nearbias is given in Srndal and Lundstrm (2005) related bias expressions, underdifferentconditions,arefoundin Bethlehem(1988) and Fuller et al. (1994). We can write alternatively nearbias(YCAL) = U ( k M k -1) yk, where M k =(U x ) k ( U k x k x) -1x . Incomparingpossiblealternatives x , a k k k convenient benchmark is the primitive auxiliary vector, x =1 for all k , which gives YCAL = U k N y r = N r y k /nr, where n is the number or r respondents, with nearbias (N yr ) = N ( yU -yU), where yU =U k yk / U and yU =U yk / N . Theratio k
relbias (Y )= CAL nearbias (Y ) CAL = nearbias( N yr )
o Samplelevel:Thevectorvalue x isknown(observed)for k every k s, and thereby known for every k r the o o unknowntotal x isestimatedwithoutbiasby sd k x . U k k
Calibrationonthiscompositeinformationcanbedonein twosteps(intermediateweightscomputedfirst,thenusedin thesecondsteptoproducefinalweights)ordirectlyinone single step. Modest differences only are expected in bias andvarianceoftheestimates.Inthesinglestepoption,the combined auxiliary vector and the corresponding infor mationare
* * x x U k k x k = o X= o . x d k x k s k
Using an extension of the instrument vector method in Section 4.3, we seek calibrated weights wk =d k vk, where vk =F ( zk) is the nonresponseadjustmentfactor, witha vector determined through the calibration equation r wk x k =X the resulting calibration estimator is YCAL = rwk yk. It is enough to specify the instrument vector value zk for respondents only zk is allowed to differfrom x . Thefunction F ( hasthesameroleasin ) k Sections4.2and4.3.Here, F ( k) implicitlyestimatesthe z inverse response probability, fk =1/ k as Deville (2002), Dupont(1995),Kott(2006) have noted. Inthelinearcase, F (u ) = 1 +u, and vk = 1 + ' zk, with ' = (U x k - r d k x k )( r d k z k x ) -1. k o The variables that make up the vector x , although k observed for sampled elements only, can be crucially important for the reduction of nonresponse bias (although * less important than the x for the reduction of variance). k For example, Beaumont (2005b) discusses data collection o process variables can be used in building the x vector k component. 9.3 Buildingtheauxiliaryvector In some surveys, there are many potential auxiliary variables,aspointedoutforexamplebyRizzo,Kaltonand Brick (1996), and Srndal and Lundstrm (2005). For example, for surveys on households and individuals in Scandinavia,asupplyofpotentialauxiliaryvariablescanbe derived from a matching of existing high quality administrative registers. A decision then has to be made whichofthesevariablesshouldbeselectedforinclusionin theauxiliaryvector x to makeitaseffectiveaspossible, k forbiasreductioninparticular.AsRizzo,KaltonandBrick (1996) point out, the choice of auxiliary variables is
( k M k - 1)yk U
N ( yU - yU )
measures how well a candidate vector x succeeds in k controlling the bias, when compared with the primitive vector. We seek an x that will give a small bias. But k relbias (Y ) isnotacomputablebiasindicatoritdepends CAL on unobserved y and on unobservable k. We need a k computableindicatorthatapproximates relbias (Y ) and CAL depends on the x vector but not on the yvariables, of whichthesurveymayhavemany. It is easy to see that relbias (YCAL) =0 if an ideal (probably nonexistent)xvectorcouldbeconstructedsuch that f k = 1/ k = xk for all k U and some constant vector . Foran xvectorthatcanactuallybeformedinthesurvey, wecanatleastobtainpredictionsofthe fk : Determine to minimize U k (fk - x ) 2 we find = , where k U -1 U =( U x k )( U k x k x ) the predicted value of f is k k fkU = U x =M k. The (thetaweighted) first and second k moment of the predictions fkU =M k are, respectively, M U = U k M k / U k = N / U = 1/ and k U
Q= 1 U k (M k - M U )2 = (1/ U )( M U - 1/ U ) U k
116
where M U =U M k / N . Srndal and Lundstrm (2007) show that relbias (Y ) and Q have under certain CAL conditionsanapproximatelylinearrelationship,
Q relbias (Y ) 1 - CAL Q0
where fU = Ufk /N and Q0 = (1/ U )( fU -1/ U ) isthe maximumvalueof Q Thusif Qwerecomputable,itcould . serveasanindicatorfor comparingthe differentcandidate x vectors. A computable analogue Q of Q is instead k obtainedasthevarianceofthecorrespondingsamplebased predictions f ks = x k =( sd k x ) s k -1 ( rd k x k x ) x =mk, sothat k k
Q= 1 rdk (mk - mr d )2 = mr d (ms d - mr d) rd k
where
mr d =
sketch of how a treatment of several of the nonresponse errorsmaybeaccomplishedunderthecaptionofcalibration thinking. Folsom and Singh (2000) present a weight calibration method using what they call the generalized exponential model (GEM). It deals with three aspects: extreme value treatment, nonresponse adjustment and calibration through poststratification.Themethodprovidesbuiltincontrolfor extreme values. Calibration to treat both coverage errors (underor overcoverage of the frame)and nonresponse is discussed in Srndal and Lundstrm (2005) and Kott (2006). Skinner (1998) discusses uses of calibration in the presence of nonresponse and measurement error. He notes something whichremainsachallengealmosttenyearslater: More research is needed to investigate the properties of calibration estimates in the presence of nonsampling errors.
rd k mk = s d k m = sd k mk . s d r dk r dk sd k
11. Conclusion
IfIamtoselectoneissueforaconcludingreflectionon the contents of this paper, let me focus on the concept of auxiliaryinformation.Itisthepivotalconceptinthepaper. If thereis notauxiliaryinformation, there is no calibration approach there is nothing to calibrate on. I noted on the other hand that regression (GREG) estimation is an alternativebutdifferentthoughtprocessforputtingauxiliary informationtoworkintheestimation. Anobjectiveinthispaperhasbeentogiveaportraitof the two types of reasoning, and I made a point of noting howthethinkingdiffers.Igaveexampleswhereessentially the same estimation objective is tackled by some authors through calibration reasoning, by others through GREG reasoning (or at least primarily by one or the other type). The respective estimators that they end up recommending may or may not agree. Whether or not the difference has significantconsequence(forvariance,forbias,forpractical matters such as consistency and transparency) depends on thesituation.Thispapermayhelpcontributinganawareness of the separation existing between two thought processes thathaveguidedresearcherssurveysampling.
Weexpectrelbiastodecreaseinaroughlylinearfashionas Q increasesthus,independentlyofthe yvariables, Q may beusedasatoolforrankingdifferentxvectorsinregardto theircapacityofreducethebias. Wecanuse Q asatooltoselect xvariablesforinclusion in the x vector, for example, by stepwise forward k selection,sothatvariablesareaddedto x oneatatime,the k variabletoenterinagivenstepbeingtheonethatgivesthe . largestincrementin Q ThemethodisdescribedinSrndal andLundstrm(2007).
10. Calibrationtoaccountforother nonsamplingerror

Nonresponse errors are critical determinants of the quality of published statistics. When we examine how the calibration approach may intervene in the treatment other sources of nonsampling error than the nonresponse, the literature to date is not surprisingly much less extensive. However, several authors sketch a calibration reasoning to also incorporate frame errors, measurement errors, and outliers. Calibration has a potential to provide a more generaltheoryforestimationinsurveys,encompassingthe variousnonsamplingerrors. As Deville (2004) points out (my translation from the French): The concept of calibration lends itself to be applied with ease and efficiency to a great variety of problemsinsurveysampling.Itsscopegoesbeyondthatof regressionestimation,anideatowhichsomeseemtowish to reduce the calibration approach. He provides a brief
References
Alexander,C.H.(1987).Aclassofmethodsforusingpersoncontrols inhouseholdweighting. SurveyMethodology,13,183198. Andersson, P.G., and Thorburn, D. (2005). An optimal calibration distance leading to the optimal regression estimator. Survey Methodology,31,9599. Ardilly, P. (2006). Les techniques de sondage. Paris: Editions Technip.
Bankier, M.D., Rathwell, S. and Majkowski, M. (1992). Two step generalizedleastsquaresestimationinthe1991CanadianCensus. Working paper, Census Operations Section, Social Surveys MethodsDivision,StatisticsCanada. Bankier,M.,Houle,A.M.andLuc,M.(1997).Calibration estimation inthe1991and1996Canadiancensuses.Proceedings, Sectionon Survey Research Methods, American Statistical Association, 66 75. Beaumont, J.F. (2005a). Calibrated imputation in surveys under a quasimodelassisted approach. Journal of the Royal Statistical Society B,67,445458. Beaumont, J.F. (2005b). On the use of data collection process informationforthetreatmentofunitnonresponsethroughweight adjustment. SurveyMethodology,31,227231. Beaumont,J.F.,andAlavi,A.(2004).Robustgeneralizedregression estimation. SurveyMethodology,30,195208. Bethlehem, J.G. (1988). Reduction of nonresponse bias through regressionestimation. JournalofOfficialStatistics,4,251260. Bethlehem,J.G., andKeller,W.J.(1987).Linearweightingofsample surveydata. JournalofOfficialStatistics,3,141153. Breidt,F.J.,andOpsomer,J.D.(2000).Localpolynomialregression estimatorsinsurveysampling.TheAnnalsofStatistics,28,1026 1053. Breidt,F.J.,Claeskens,G.andOpsomer,J.D.(2005).Modelassisted estimation for complex surveys using penalized splines. Biometrika,92,831846. Chambers, R.L. (1996). Robust caseweighting for multipurpose establishmentsurveys. JournalofOfficialStatistics,12,332. Chambers,R.L., Dorfman,A.H.andWehrly,T.E.(1993).Biasrobust estimationinfinitepopulationsnonparametriccalibration. Journal oftheAmericanStatisticalAssociation,88,268277. Deming, W.E. (1943). Statistical Adjustment of Data. New York: John Wiley &Sons, Inc. Demnati, A., and Rao, J.N.K. (2004). Linearization variance estimatorsforsurveydata. SurveyMethodology,30,1726. Deville,J.C.(1998).Lacorrectiondelanonrponseparcalageoupar chantillonnage quilibr. Paper presented at the Congrs de lACFAS,Sherbrooke,Qubec. Deville, J.C. (2002). La correction de la nonrponse par calage gnralis. ActesdesJournesdeMthodologie,I.N.S.E.E.,Paris. Deville,J.C.(2004).Calage,calagegnralisethypercalage. Internal document,INSEE,Paris. Deville, J.C., and Srndal, C.E. (1992). Calibration estimators in surveysampling.JournaloftheAmericanStatisticalAssociation, 87,376382. Deville, J.C., Srndal, C.E. and Sautory, O. (1993). Generalized raking procedures in survey sampling. Journal of the American StatisticalAssociation,88,10131020. Duchesne, P. (1999). Robust calibration estimators. Survey Methodology,25,4356. Dupont, F. (1995). Alternative adjustments where there are several levelsofauxiliaryinformation. SurveyMethodology,21,125135. Ekholm, A., and Laaksonen, S. (1991). Weighting via response modeling in the Finish Household Budget Survey. Journal of OfficialStatistics,3,325337. Estevao, V.M., and Srndal, C.E. (2000). A functional form approachto calibration. JournalofOfficialStatistics,16,379399.
117
Estevao,V.M.,andSrndal,C.E.(2002).Thetencasesofauxiliary information for calibration in twophase sampling. Journal of OfficialStatistics,18,233255. Estevao,V.M.,andSrndal,C.E.(2004).Borrowingstrengthisnot thebesttechniquewithinawideclassofdesignconsistentdomain estimators. JournalofOfficialStatistics,20,645660. Estevao, V.M., and Srndal, C.E. (2006). Survey estimates by calibration on complex auxiliary information. International StatisticalReview,74,127147. Firth, D., and Bennett, K.E. (1998). Robust models in probability sampling. JournaloftheRoyalStatisticalSociety,60,321. Folsom, R.E.,and Singh, A.C. (2000). The generalized exponential model for design weight calibration for extreme values, nonresponse and poststratification. Proceedings, Section on SurveyResearchMethods,AmericanStatisticalAssociation,598 603. Fuller, W.A. (2002). Regression estimation for survey samples. Survey Methodology,28,523. Fuller, W.A., Loughin, M.M. and Baker, H.D. (1994). Regression weightinginthepresenceofnonresponsewithapplicationtothe 19871988 nationwide Food Consumption Survey. Survey Methodology,20,7585. Fuller, W.A., and Rao, J.N.K. (2001). A regression composite estimatorwithapplicationtotheCanadianLabourForceSurvey. SurveyMethodology,27,4551. Hansen,M.H., andHurwitz,W.N.(1943).Onthetheoryofsampling from finite populations. Annals of Mathematical Statistics, 14, 333362. Harms,T.(2003).Extensionsofthecalibrationapproach:Calibration of distribution functions and its link to small area estimators. Chintex working paper no. 13, Federal Statistical Office, Germany. Harms, T., and Duchesne, P. (2006). On calibration estimation for quantiles. SurveyMethodology,32,3752. Hidiroglou,M.A.(2001).Doublesampling. SurveyMethodology,27, 143154. Hidiroglou, M.A., and Srndal, C.E. (1998). Use of auxiliary informationfortwophasesampling. SurveyMethodology,24,11 20. Horvitz, D.G., and Thompson, D.J. (1952). A generalization of sampling without replacement from a finite universe.Journal of theAmericanStatisticalAssociation,47,663685. Huang, E.T., and Fuller, W.A. (1978). Nonnegative regression estimation for sample survey data. Proceedings,Social Statistics Section,AmericanStatisticalAssociation,300305. Isaki, C.T., Tsay, J.H. and Fuller, W.A. (2004). Weighting sample datasubjecttoindependentcontrols. SurveyMethodology,30,35 44. Kalton, G.,and FloresCervantes, I. (1998). Weighting methods. In New Methods for Survey Research (Eds. A.Westlake, J. Martin, M.RiggandC.Skinner),.Berkeley,U.K.:AssociationforSurvey Computing. Kim, J., Breidt, F.J. and Opsomer, J.D. (2005). Nonparametric regression estimation of finite population totals under twostage sampling.Unpublishedmanuscript. Knottnerus, P., and van Duin, C. (2006). Variances in repeated weightingwithanapplicationtotheDutchLabourForceSurvey. JournalofOfficialStatistics,22,565584.
118
Plikusas, A. (2006). Nonlinear calibration. Proceedings, Workshop on Survey Sampling, Venspils, Latvia. Riga: Central Statistical BureauofLatvia. Ren,R.(2002).Estimationdelafonctionderpartitionetdesfractiles dune population finie. Actes des journes de mthodologie statistique,INSEEMthodes,tome1,100,263289. Renssen, R.H., and Nieuwenbroek, N.J. (1997). Aligning estimates forcommonvariablesintwoormoresamplesurveys.Journalof theAmericanStatisticalAssociation,92,368374. Renssen,R.H.,Kroese,A.H.andWilleboordse,A.J.(2001).Aligning estimates by repeated weighting. Report, Central Bureau of Statistics,TheNetherlands. Rizzo,L.,Kalton,G.andBrick,J.M.(1996).Acomparisonofsome weighting adjustment methods for panel nonresponse. Survey Methodology,22,4353. Rueda, M., Martnez, S., Martnez, H. and Arcos, A. (2007). Estimation of the distribution function with calibration methods. JournalofStatisticalPlanningandInference,137,435448. Srndal,C.E.(1982). Implications ofsurveydesign for generalized regression estimation of linear functions. Journal of Statistical PlanningandInference,7,155170. Srndal, C.E., and Swensson, B. (1987). A general view of estimation for two phases of selection with application to two phasesamplingandnonresponse. InternationalStatisticalReview, 55,279294. Srndal,C.E.,Swensson,B.andWretman,J.(1992).Modelassisted SurveySampling.NewYork:SpringerVerlag. Srndal,C.E.,andLundstrm,S.(2005).EstimationinSurveyswith Nonresponse.NewYork: John Wiley &Sons,Inc. Srndal,C.E., andLundstrm,S.(2007).Assessingauxiliaryvectors for control of nonresponse bias in the calibration estimator. Statistics Sweden: Research and development methodology report2007:2,toappear, JournalofOfficialStatistics. Singh, A.C., and Mohl, C.A. (1996). Understanding calibration estimatorsinsurveysampling. SurveyMethodology,22,107115. Singh, S., Horn. S. and Yu, F. (1998). Estimation of variance of general regression estimator: Higher level calibration approach. SurveyMethodology,24,4150. Skinner, C. (1998). Calibration weighting and nonsampling errors. Proceedings International Seminar on New Techniques for Statistics, Sorrento,November46,1998,5562. StatisticsCanada(1998).MethodologyoftheCanadianLabourForce Survey. Statistics Canada, Household Survey Methods Division. Ottawa:MinisterofIndustry,catalogueno.71526XPB. StatisticsCanada(2003).QualityGuidelines(fourthedition).Ottawa: MinisterofIndustry,Catalogueno.12539XIE. Steel,D.G., andClark,R.G.(2007).Personlevelandhouseholdlevel regressionestimationinhouseholdsurveys.SurveyMethodology, 33,5160. Stukel,D.M.,Hidiroglou,M.A.andSrndal,C.E.(1996).Variance estimationforcalibrationestimators:Acomparisonofjackknifing versusTaylorlinearization. SurveyMethodology,22,117125. Thberge, A. (1999). Extension of calibration estimators in survey sampling. Journal of the American Statistical Association, 94, 635644. Thberge, A. (2000). Calibration and restricted weights. Survey Methodology,26,99107.
Kott, P.S. (2004). Comment on Demnati and Rao: Linearization varianceestimatorsforsurveydata.SurveyMethodology,30,27 28. Kott, P.S. (2006). Using calibration weighting to adjust for nonresponse and coverage errors.Survey Methodology, 32, 133 142. Kovaevi, M.S. (1997). Calibration estimation of cumulative distributionandquantilefunctionsfromsurveydata.Proceedings oftheSurveyMethodsSection,StatisticalSocietyofCanada,139 144. Krapavickait,D., and Plikusas,A.(2005).Estimationofaratiointhe finitepopulation. Informatica,16,347364. Lavalle,P.(2007). IndirectSampling.NewYork:SpringerVerlag. LeGuennec, J., and Sautory, O. (2002). CALMAR2: une nouvelle versiondelamacroCALMARderedressementdchantillonpar calage. ActesdesJournesdeMthodologie,INSEE,Paris. Lehtonen, R., and Veijanen, A. (1998). Logistic generalized regressionestimators. SurveyMethodology,24,5155. Lehtonen,R.,Srndal,C.E.and Veijanen, A.(2003).Theeffect of modelchoiceinestimationfordomains,includingsmalldomains. SurveyMethodology,29,3344. Lehtonen,R.,Srndal,C.E.andVeijanen,A.(2005).Doesthemodel matter? Comparing modelassisted and modeldependent estimators of class frequencies for domains. Statistics in Transition,7,649674. Lematre, G., and Dufour, J. (1987). An integrated method for weighting persons and families. Survey Methodology, 13, 199 207. Lundstrm, S.,and Srndal, C.E. (1999). Calibration as a standard method for treatment of nonresponse. Journal of Official Statistics,15,305327. Merkouris, T. (2004). Combining independent regression estimators from multiple surveys. Journal of the American Statistical Association,99,11311139. Montanari, G.E. (1987). Postsampling efficient prediction in large scalesurveys. InternationalStatisticalReview,55,191202. Montanari, G.E.,and Ranalli, M.G. (2003). On calibration methods for designbased finite population inferences. Bulletin of the th International Statistical Institute, 54 session, volume LX, contributedpapers,book2,8182. Montanari, G.E., and Ranalli, M.G. (2005). Nonparametric model calibration estimation in survey sampling. Journal of the AmericanStatisticalAssociation,100,14291442. Myrskyl, M. (2007). Generalised regression estimation for domain classfrequencies.StatisticsFinlandResearchReports247. Nieuwenbroek, N.J. (1993). An integrated method for weighting characteristics of persons and households using the linear regression estimator. Report, Central Bureau of Statistics, The Netherlands. Nieuwenbroek, N.J., and Boonstra, H.J. (2002). Bascula 4.0 for weighting sample survey data with estimation of variances. The SurveyStatistician,SoftwareReviews,July2002. Nieuwenbroek,N.J., Renssen,R.H.andHofman,L.(2000).Towards a generalized weighting system. In Proceedings, Second International Conference on Establishment Surveys, American StatisticalAssociation,AlexandriaVA. Park, M. and Fuller, W.A. (2005). Towards nonnegative regression weightsforsurveysamples. SurveyMethodology,31,8593.
Till,Y.(2002).Unbiasedestimationbycalibrationondistributionin simple sampling designs without replacement. Survey Methodology,28,7785. Tracy, D.S.,Singh,S.and Arnab,R.(2003).Note on calibrationin stratifiedanddoublesampling. SurveyMethodology,29,99104. Vanderhoeft,C.(2001).GeneralisedcalibrationatStatisticsBelgium. SPSS Module gCALIBS and current practices. Statistics Belgium Working Paper no. 3. Available at: www.statbel.fgov.be/studies/paper03_en.asp. Vanderhoeft,C.,Waeytens,E.andMuseux,J.M.(2001).Generalised calibration with SPSS 9.0 for Windows baser. In Enqutes, Modles et Applications (Eds. J.J. Droesbeke and L. Lebart), Paris:Dunod
119
Webber, M., Latouche, M. and Rancourt, E. (2000). Harmonised calibration of income statistics. Statistics Canada, internal document,April2000. Wu, C. (2003). Optimal calibration estimators in survey sampling. Biometrika,90,937951. Wu, C., and Sitter, R.R. (2001). A modelcalibration approach to usingcompleteauxiliaryinformationfromsurveydata. Journalof theAmericanStatisticalAssociation,96,185193. Zheng, H., and Little, R.J.A. (2003). Penalized spline modelbased estimation of the finite population total from probability proportionaltosizesamples. JournalofOfficialStatistics,19,99 117. Zieschang,K.D.(1990).Sampleweightingmethodsandestimationof totals in the Consumer Expenditure Survey. Journal of the AmericanStatisticalAssociation,85,9861001.

Sarndal (2007)

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sarndal (2007)

Uploaded by

Copyright:

Available Formats

Component of Statistics Canada Catalogue no.

12-001-X Business Survey Methods Division

Article The calibration approach in survey theory and practice

SurveyMethodology,December2007 Vol.33,No.2,pp.99119 StatisticsCanada,Catalogueno.12001X

Gk ( wk , dk ) subject to the calibration equation s s wk xk =U x leads to wk =d k F ( qk x ), where is k k obtainedasthesolution(assumingoneexists)of

Forsimplicity,letustake qk =1 forall kwederivethe calibratedweights,rearrangetermsandfindthatthemodel calibrationestimatorcanbewrittenas

where ysd = s d k yk / s d k ysd = s d k yk / sd k, and % Bs d =

) Acalibrationestimator Fy(t oftakestheform FyCAL(t ) = 1 wk D(t - yk) swk s (6.2)

swk ek = 0. swk x2k

where xk isthecombinedauxiliaryvector x x = 1k . k 2 x k Alternatively, in asingle step option, we determine the w directlytosatisfy k

10. Calibrationtoaccountforother nonsamplingerror

You might also like