The dilemma of survey nonresponse(6)

来源：网络收集时间：2026-09-15

导读： Whensamplingbiasisidentifiedusingthemethodsoutlinedpreviously,reweightingofobtaineddataisthestrategymostfrequentlyusedtocompensateforsamplingbiascausedbynonresponse.Usingastatisticalpackage,suchasSPS

Whensamplingbiasisidentifiedusingthemethodsoutlinedpreviously,reweightingofobtaineddataisthestrategymostfrequentlyusedtocompensateforsamplingbiascausedbynonresponse.Usingastatisticalpackage,suchasSPSS(SPSS,Inc,Chicago,IL),selectedcasescanbegivenadditionalweightinanalyses.Theweightingfactorsarechosentoremedythedifferencebetweensampleproportionsandpopulationproportionsforvarioussubgroups.Forexample,ifthepopulationissplitevenlybetweenmenandwomen,but60%oftherespondentsaremen,weightscanbeappliedtocompensateforthedifference.Theweightappliedtoeachrespondentiscalculatedasthepopulationproportiondividedbythesampleproportion.Inthiscase,theweightformalerespondentswouldbe50%dividedby60%,or0.8333.Theweightappliedtofemalerespondentswouldbe50%dividedby40%,or1.25.Whenstatisticssuchasaveragesarecalculatedovertheweighteddata,itisasifthesamplewassplit50/50intermsofgender.Holt,Elliott,andMoore(1999),usedanalternative,mathematicallyequivalentmethodforcalculatingweights.Theycalculatedtheweightofeachcaseasthenumberintheweightingclassfortheentirepopulationdividedbythenumberintheweightingclassforthesample.Foramoredetaileddiscussionofweightingandotherstatisticaltechniquestocompensatefornonresponse,seeGroves(1989),Grovesetal.(2001),LittleandRubin(1987),orKalton(1983).

Althoughweightingiswidelyusedtocompensatefornonresponse,weightingadjustmentsarenotalwayseffectiveinreducingbias(Brehm,1993;vanGoor&Stuiver,1998),and,insomecases,theymayevenexacerbatetheproblem(Brehm,1993).Essentially,theprocessof

250J.Burkell/Library&InformationScienceResearch25(2003)239–263

weightingreplicatesrespondentswithineachweightingclasstocompensateforthosewithintheclasswhodidnotrespond.Thus,weightingisbasedontheassumptionthat,withineachweightingclass,therearenosystematicdifferencesbetweenrespondentsandnonrespondents(Mandell,1974).Ifthisassumptionisnotmet(andwithoutactuallycollectingdatafromnonrespondents,thereisnowaytodeterminewhetheritis),weightingcannotcompensateforthesamplingbiasintroducedbynonresponse.Morecomplexmodelingapproaches(Achen,1986;Glynn,Laird,&Rubin,1986;Heckman,1976,1979;Heckman&Robb,1986)arelesssusceptibletothiscriticism,althoughtheseapproachesrequirethattheresearchermakeassumptionsabouttherelationshipbetweenthelikelihoodofparticipationandthevariablesofinterest(Brehm,1993;Wainer,1989).Deming(1960)suggestedtheonlyapproachthatdoesnotrequireassumptionsaboutthenatureofeitherthemechanismsthatresultinnonresponseorthedatathatwouldhavebeenobtainedfromnonrespondents.Inthatapproach,thedatafromtheconvertednonrespondentsrepresentthedatafromtheentiregroupthatdidnotrespond.Iftheresearcherissuccessfulincollectingdatafromarandomsubsetoftheoriginalnonrespondents,appropriateweightingofthedataprovidedbythisgroupadequatelycorrectsfornonresponsebias.

4.NonresponseintheLISliterature

Summariesofsurveyspublishedinavariousdisciplinesindicatethat,althoughresponseratesarelowenoughtoconstituteanissuefortheinterpretationofresults,alargeproportionofpublishedarticlesfailtorespondtotheissue(Cummingsetal.,2001;Lindner,Murphy,&Briers,2001).ThissectionattemptstodeterminewhetherasimilarsituationexistsintheLISliterature.Specifically,anempiricalstudyofsurveyspublishedinLISisusedtoexaminethefollowingquestions:

??????

WhatistheextentofnonresponseinsurveyspublishedintheLISliteratureandhowseriousisthislevelofnonresponsefortheinterpretationofsurveyresults?Whatmethodologicalstrategiesdoresearcherstaketoreducenonresponse?

Doresearchersattempttodeterminethedegreeofand/orstatisticallyaccountfornonresponsebias,andifso,how?

Conclusionsarebasedonacensusofsurveyspublishedfrom1996through2001inthreeLISjournals:PublicLibraries,College&ResearchLibraries,andLibrary&InformationScienceResearch.Theyears1996through2001representthemostrecentsixfullyearsofpublicationforthesejournalsatthetimethecensuswascompleted.Thisconstitutesapurposivesample,chosentocoverabroadrangeofrecentLISresearchinbothacademicandpubliclibrarysettings.ThesurveysexaminedarenotarandomsampleofsurveyspublishedinLIS,andthereforecaremustbetakeningeneralizingbeyondthespecificresults.Thepurpose,however,isonlytodeterminewhethernonresponseandthetreatmentofitareaproblemintheLISliteratureandnottodrawquantitativeconclusionsaboutthedegreeoftheproblem.Thepurposivesampledescribedissufficient,ifnotideal,tomeetthoseneeds.

J.Burkell/Library&InformationScienceResearch25(2003)239–263251

Thesurveysselectedmetthefollowingcriteria:

??????

Thegoalofthesurveywastodescribethepopulationofinterest(thus,surveyswerenotincludediftherewasnointentiontoapplyresultstothepopulation).

Eithertheentirepopulationwasselectedforparticipation(acensus)ortheselectedparticipantsrepresentedarandomsampleofthepopulation.Theresponseratewasreported.

TheAppendixpresentsthelistofarticlesthatreportedthesurveysusedintheanalysis.Fromthepublishedreportofeachsurvey,informationwasextractedregardingtheresponserate,themethodologicalstrategiesusedtoreducenonresponse,andanyattemptstoassessand/orcorrectforbiascausedbynonresponse(seeTable3fordetails).Inmanyofthereportsexamined,themethodologicaldetailsaresketchy(thisisconsistentwiththeresultsofMcKechnieetal.,2002).Forthepurposesofthisanalysis,itisassumedthatmethodologicalvariationsarenotusedunlesstheyareexplicitlymentionedinthereport.Anybiasthatarisesfromthisassumptionwillleadtotheunderestimationoftheuseofthestrategies.Detailssuchasthese,however,shouldbeincludedinthemethodsdescriptionsofpublishedstudies,andthereforetheassumptionisdeemedappropriate.

Table3

RecordedsurveyinformationVariableResponserateGroupsurveyed

LevelofmeasurementRatioNominal