The dilemma of survey nonresponse(4)

First,theimpactofnonresponseonsurveyresultsdependsentirelyontherelationshipbetweenthemechanismsthatresultinnonresponseandthevariablesofinterest.Again,bydefinition,theresearcherisnotinapositiontomeasurethisrelationshipbecausetherequired

246J.Burkell/Library&InformationScienceResearch25(2003)239–263

dataareunavailable.Itispossible,however,tohypothesizeabouttherelationship.Totheextentthattheresearchercanmakeastrong,logicalcasethatthelikelihoodofresponseisunrelatedtosurveyconclusions,thereissomejustificationforignoringnonresponse.

Second,thedegreeofnonresponseisimportant.Potentialbiasincreasesmonotonicallywithnonresponse(Alexanderetal.,1984;Chen,1996;Cochran,1963;vanGoor&Stuiver,1998;Viswesvaranetal.,1993):thehigherthelevelofnonresponse,thegreaterthepotentialbias.Giventhatanydegreeofnonresponsecanleadtobiasinsurveyresults,itisdifficult,ifnotimpossible,toidentifyanacceptablelevelofresponse.Nonetheless,itiswidelyheldthataresponserateof75%to90%issufficienttosupportgeneralization.Iftheresponserateforasurveyismorethan90%,researchersareprobablyjustifiedingeneralizingfromthesampletothepopulation.Iftheresponseratefallsbetween75%and90%,generalizationmaybejustified,particularlyifotherconditionsidentifiedinthissectionarealsomet.Whensurveyresponseratesfallbelow75%,straightforwardgeneralizationsfromsampletopopulationaretenuousatbest,andthedatashouldprobablybesubjectedtosomeoftheotherproceduresoutlinedinthissectiontoassessandpossiblycorrectforbiascausedbynonresponse.

Third,thetypeofresearchisimportant.Nonresponseislessseriousforpreliminaryresearchorresearchdesignedtosupporttheorydevelopment—inpartbecausesuchstudiesarerarelyinterpretedalone.Inaddition,manystudiesofthistypefocusonrelationshipsbetweenmultiplevariables(discussedlater)thatmaybelessinfluencedbynonresponse.Finally,thereissomeevidencethattheimpactofnonresponseisgreaterfortheestimationofunivariatepopulationparameters(e.g.,averageageoflibraryusers)thanfortheestimationofbivariatecorrelations(e.g.,relationshipbetweenageandlibraryresourceuse),multivariaterelationships(e.g.,facultystatus,salary,andlibrarysizeaspredictorsoflibrarianjobsatisfaction),ortestsofdifferencesbetweengroupswithinthelargersample(e.g.,comparisonoftheeffectivenessofbibliographicinstructionforface-to-faceversusdistancelearners).Caremustbetaken,however,intheapplicationofthiscriterion.Althoughasmallerimpactofnonresponseonbivariateandunivariaterelationshipshasbeenobservedinempiricaldata(Goodman&Blum,1996;vanGoor&Verhage,1999)andthereissometheoreticalbasistoassumethattheobservationiscorrect(Alexander,Barrett,Alliger,&Carson,1986),thisconclusionisnotunchallenged.Brehm(1993)presentedconvincingtheoreticalevidenceandsupportingMonteCarlosimulationdatademonstratingthatnonresponsecanhavealargebiasingeffectonmultivariaterelationshipsiftheselectioncriteriaandthedependentvariableofinterestarerelated.

3.3.Limitingsurveyconclusions

Thepreviousdiscussionindicatesthat,inmanycasesifnotmost,itisinappropriatesimplytoignorenonresponse.Analternativestrategyistolimitsurveyconclusionscommensuratewiththelevelofnonresponse.Ifthesurveyresponserateislessthan75%,theresearchershould,ataminimum,ensurethatthereaderisawareofthelimitationtothegeneralizabilityofsurveyresultsarisingfromnonresponse.Abetterapproachmaybetoassumethatnonresponsehasintroducedbiasinsurveyresults;itisthenpossibletoestimatethelikelyimpactofthatbiasusingstatisticaltechniques.Thesetechniquesworkbycalculatinghow

J.Burkell/Library&InformationScienceResearch25(2003)239–263247

nonrespondentsmighthaveansweredusingworstcaseoraveragecaseassumptions;researchersthenusethisinformationtoreducetheprecisionwithwhichresultsarestated.Thereareproceduresavailableforcalculatingthemaximumamountofbiasintroducedbynonresponseinaunivariatepopulationestimate,suchastheestimateofaveragejobsatisfactionamonglibrarians(Alexanderetal.,1984;Chen,1996;Cochran,1963).Somesurveyanalysesincludesubgroupcomparisons,suchasanexaminationofthecomplianceofsmall,medium,andlargepubliclibrarieswithrecommendationsregardingInternetconnec-tivity.Whensubgroupdifferencesaretestedusingttestsoranalysesofvariance,itispossibletoestimatethedegreetowhichsuchdifferencesareaffectedbynonresponse(Viswesvaranetal.,1993).Thesetechniqueseffectivelyreducetheprecisionofsurveyresultstocompensateforsamplingbiasfromnonresponse.Unlessresponseratesareveryhigh,however,theytendtobeoverlyconservative,andresearcherswhousetheseapproachesarelikelytounderes-timatethesignificanceoftheirresults.Thus,thesestatisticsarenotwidelyusedbecausetheyworkbestwhentheyareneededleast:atlowlevelsofnonresponse.3.4.Assessingandcorrectingforsamplingbiascausedbynonresponse

Amorewidelyusedapproachtotheissueofnonresponseistoattempttodeterminewhethernonresponsehasintroducedsamplingbiasandtocorrectthatbiasifitisfound.Themostcommonlyusedstrategycomparesrespondentstoasecondgroupthateitherincludesorrepresentsnonrespondentsusingoneofthethreespecificstrategiespresentedinthissection.Qualitativecomparisons(e.g.,Clougherty,Forys,Lyles,Persson,Walters,&Washington-Hoagland,1998;Hart,1998)providesomeinsightintotheimpactofnonresponse,butabetterapproachistotestthestatisticalsignificanceofanyobserveddifferences.Asignificantdifferencebetweenrespondentsandthecomparisongroupsuggeststhepresenceofsamplingbias.Ifsuchadifferenceisfound,researcherscanattempttocorrectthebiasbyreweightingthedatatoaccountfortheunderrepresentationofsubgroupsresultingfromthenonresponse(seesection3.5).

Therespondentscanbecomparedwiththepopulationfromwhichtheyweredrawninanattempttoanswerthefollowingquestion:Dorespondentsrepresentarandomsampleofthepopulation?Theapplicationofthisapproachislimitedtocomparisonsforwhichdataareavailablebothforrespondentsandforthepopulationasawhole—generallydemographicdescriptors.Astatisticaltestisappliedtothesedatatodeterminetheprobabilitythatthesampleofrespondentsisdrawnfromthepopulation.Theappropriatestatisticaltestdependsonthelevelofmeasurementofvariablebeingcompared(intervalorratioversusord …… 此处隐藏：5773字，全部文档内容请下载后查看。喜欢就下载吧 ……