Vol. 9(8), pp. 208-215, 23 April, 2014DOI: 10.5897/ ERR2014.1709 ISSN 1990-3839Copyright 2014 Author(s) retain the copyright of this article http:/ / www.academicjournals.org/ ERR Educational Research and Reviews Full Length Research Paper The effect of misssing data handling methods on goodness of fit indices in confirmatory factor analysis Alper Kse Abant Izzet Baysal University-Turkey. Recieved 1st March 2014; Accepted 15th April 2014 Theprimaryobjectiveofthisstudywastoexaminetheeffectofmissingdataongoodnessoffit statisticsinconfirmatoryfactoranalysis(CFA).Forthisaim,fourmissingdatahandlingmethods; listwisedeletion,fullinformationmaximumlikelihood,regressionimputationandexpectation maximization (EM) imputation were examined in terms of sample size and proportion of missing data. It isevidentfromtheresults that whentheproportionsofmissingness%1orless,listwisedeletioncan bepreferred.Formoreproportionsofmissingness,fullinformationmaximumlikelihood(FIML) imputationmethodshowsvisibleperformanceandgivesclosestfitindicestooriginalfitindices.For this reason, FIML imputation method can be preferred in CFA. Key words:Missing data, goodness of fit, confirmatory factor analysis, incomplate data, and missing value.

INTRODUCTION Educationalandpsychologicalscientistshaveimproved their ability to carry out quantitative analysis on large and complex data bases with the use of computers. The goal ofthe researcheris runningthemost preciseanalysisof thedataformakingacceptableandeffectivedeductions aboutthepopulation(SchaferandGarham,2002).Generally, scientists have ignored or have underestimated somekindofresearchproblemsbyreasonofmissing databutwiththehelpofimprovedtechnologyand computers, this problem can be handled easily. The term missingdatameansthatsometypeofinterested informationaboutthephenomenaismissing(Kenny, 2005). Missing data is one of the most common problems in data analysis.Theproblemoccursasaresultof various factors. Equipment errors, reluctant respondents, researchergoofscanbegivenasanexampleforthese factors.Quantityandpatternofmissingdatadetermines its seriousness for the research (Tabachnick andFidell, 2001). Researchesonmissingdatadatesasfarbackasthe 1930s.Atfirst,amaximumlikelihoodmethodforimpu-tationofmissingdatainbivariatenormaldistributions wassuggestedbyWilks(1932:citedinCheema,2012). Severalmethodssuchaslinearregressionwereintro-duced between the 1950s and 1960s. However, absence ofstatisticalsoftwarescausedlittleprogressforhowto handlemissingdataatthattime.In1980sand1990s, developedcomputer packagesmadehandlingmising

E-mail: i.alper.kose@gmail.com Author agree that this article remain permanently open access under the terms of the Creative Commons Attribution License 4.0 International License data easy for resear-chers (Cheema, 2012).Itispossibletofacetomissingdatainbehavioral sciences.Inallresearchstudies,reportingnecessary information for missing data should be done.Resarchers shouldreporttheextentandnatureofmissingdataand theproceduresforhowtohandlethemissingdata (Schlomeretal.,2010).Academicjournalsexpectfrom theauthorstotakeappropriatestepstoproperlyhandle missingdata,butmostarticlesdonotgivenecessary attention to this isue (Sterner, 2011).Small percentages ofmissingvaluesarelessproblematicbutthereisno commondefinitionofsmallamountofmissingdatain the literature (Saunders et al., 2006).Researchersshouldalsotakeintoconsiderationthe pattern of missing data as well as amount and source of missingdata.Occasionally,patternofmissingdatacan becalledasmissingdatamechanism.Itisessentialto remindherethatthewordmechanismisusedasa technical term. It gives point to structural association with the missing data and the observed and/or missing values ofothervariablesinthedatawithoutemphasizingthe hypotheticalprimaryreasonoftheseassoications (Kenny,2005).Missingcompletelyatrandom(MCAR), missingatrandom(MAR),andnotmissingatrandom (MNAR)arethreepatternsofmissingness(Schlomeret al., 2010; Cheema, 2012). Missing completely at random Missingvaluesdonothaveanyrelationshipwithany variablebeingexaminedandmissingvaluesrandomly distributed throughout data. In other words, probability of missing on Y does not depend on neither X or Y (Schafer andGarham,2002;Acock,2005;Schlomeretal.,2010; Sterner, 2011). Mathematically this can be presented as . For example, a student does not finish a test because of his or her instantaneous health problem. Missing at random WithMAR,missingdatamayberelatedtoatleastone variableinthestudybutnottotheoutcomebeing measured(SchaferandGarham,2002).Itcanbestated as . Forexample,itcouldbedifficultforanelderlypersonto finish the questionnaire byreasonofage(ameasured Kse209 variable) but not because of his or her level of depression (the outcome being measured) (Saunders et al., 2006). Missing not at random With MNAR, the reason for missingness is related to one or more of the outcome variable or the missingness has a systematicpattern(SchaferandGarham,2002).In mathematical base, it can be formulated as; ). Itmeansthatsomethingwhichyouhavenotmeasured asandeterminativefactorforthepossibilitythatan observationismissing(DaveyandSalva,2010; MolenberghsandKenward,2007).Forexample,ifthe participantsdontthinkthereisaprogressfromthe treatment,theymaygiveupastudyondepressionand not complete the final questionnaire. Statisticalresultscanbemoreaffectedbythepattern of missing data than the percentage of missingness. The patternofmissingdatabasiclybasedonrandomnessof missingvalues.Randomnessislessproblematicthan nonrandomness.Becausenonrandomnessaffectsthe generalizability of results (Tabachnick and Fidell, 2001). Allmeasures,lessormore,containsomemeasure-menterrorinevitably.Statisticalanalysisresultsand conclusionsdrawnfromtheseresultsareaffectedby measurementerrors.Missingdataresulttoeithermea-surementerrororsamplingerror,dependingonhowthe missingdataarehandled(Mackelprang,1970).Resear-cherscanhandlemissingdatawiththehelpofvarious methodswhichhavedifferenteffectsonestimationand anddecisionsmadeonthebasisoftheseestimations. There are various methods which are available to handle themissingdataproblem.Thesemethodscanbe classifiedasoldmethodsandnewmethods.Old methodsrequirelessmathematicalcomputations.In contrasttooldmethodsnewmethodsbasedonmore complexmathematicalcomputations.(Saundersetal., 2006). Deletion Methods Listwisedeletion:Listwisedeletionalsocalledasa completecaseanalysisisthemethodwhichisthemost wide spread and easiest of all to handle the missind data (Schafer and Garham, 2002; Acock, 2005; Enders, 2001; Enders,2013).Computerprogramautomatically dischargesmissingcasesfromthedatawhenlistwise deletion is used. This procedure reduces sample size and itcausesstatisticalpowerreductionandresearchers should ask the representativeness of remaining sample.210 Educ. Res. Rev. Samplesizereductionbringsbiasproblem,infliationof standart errors and reduction of significane level. (Acock, 2005; Saunders et al., 2006). Pairwisedeletion:Pairwisedeletionisverysimilarto listwisedeletionbutthismethoddischargesmissing cases only in the analysis. For instance, if there are three variablesasX,YandZandmissingcaseisonZ variable.Correlationwilluseallnobservationsto calculaterXYbutonlyn-1observationstocalculaterXZ and rYZ (Cheema, 2012). Imputation methods Meansubstitution:Thisiseasyandfastmethodin whichallmissingcasesissubstitutedbythemeanof totalsample(Saundersetal.,2006).Thismethodhas somedisadvantagespractically.Usageforrespondents attheextremescancausemisleadingresults.Richand poorpersonswouldnotwanttogivetheirincomesina telephone survey. If mean of the population is substituted for this maissing part, it would be spurious guess (Acock, 2005).Meansubstitutiondoesntchangevariablemean butitcanbeusedonlyifthemissingpatternisMCAR. But it has reducing effect on variance and causes biased anddeflatederrors(Pigott,2001;TabachnickandFidell, 2001). Regressionimputationorconditionalmean imputation:Regressionimputationismorecomplicated methodandbasedonregressionequations.Selected predictors (highest correlations) are used as independent andmissingdataisusedasdependentvariables.With the help of repeated regression equations missing values are predicted (Saunders et al., 2006; Peugh and Enders, 2004). This methods advantage is objectiveness against researchers guess. Two disadvantages can be given for this method. First, predictions from other variables for the regressionequationscausebeterfitthantherealscore. Second,itcausesreducingvariance(Tabachnickand Fidell,2001).Thesemethodsareknownastheconven-tionalmethodsintheliteratureandtheyproducebiased estimates of parameters or their standard errors. EM and multipleimputationarenewmethodsthathavemuch better statistical properties (Allison, 2003). Maximumlikelihood(Ml):MLisknownasmodern methodthatutilizesinformationfromothervariables duringparameterestimationprocedurebyincorporating informationfromtheconditionaldistributionofobserved variables. There are three maximum likelihood estimation algorithims;themultiplegroupapproach,fullinformation maximumlikelihood(FIML)estimationandexpectation maximizationalgorithim.Allthreealgorithimsassume multivariate normality (Enders, 2001).The multi-group approach is difficulttoimplementand stipulatesanexceptionallevelofexpertise.Forthis reasonthisapproachisnotwidespreadamongresear-chers but the multiple group approach can be conducted in all structural equation modeling (SEM) softwares. FIML canbeseenassimilartothemultiple-groupmethodbut likelihoodfunctioniscalculatedattheindividuallevel, rather than the group level (Enders, 2001). Amos and Mx offer FIML. FIML was recommended as a superior method fordealingwithmissingdatainstructuralequation modeling.Specifically,EndersandBandalos(2001) pointedoutthatFIMLestimatesareunbiasedandeffi-cient under MCAR and MAR mechanisims. The third ML algorithimisexpectation-maximazationalgorithim.EM estimatesmissingdatavaluesonthelikelihoodunder that distribution. EM is an iterative procedure and includec two steps; expectation (E) and maximization (M) for each iteration(TabachnickandFidell,2001).WithEstep,the conditionalexpectationof theparameteris calculatedon missingdata.Itisconductedbyaseriesofregression equations (Enders, 2001). With M step the parameters by maximizingthecompletedatalikelihoodareestimated (Jing,2012).Statisticalpackageforthesocialsciences (SPSS),estimationofmeansandcovariances(EMCOV) andNORMofferEMalgorithm.Becauseofthe accessibility for SPSS, EM algorithim was selected. Multipleimputation:MIisananothercomplicated missingdatahandlingmethodinwhichanumberof imputeddatasets(frequentlybetween5and10)are createdandinthesedatasetsdifferentestimationof missingvaluesareavailable.Theseparameterestima-tions for missing values are avareged to produce a single setofresults(PeughandEnders,2004;Roseand Fraser,2008).Researchershavebeenrecommended maximumlikelihoodandmultipleimputationmethod becauseoftheirlessrestrictiveassumptions,strong theorethical assumptions, less biased results and greater statisticalpower(Enders,2013).EndersandBandalos (2001)statedthatmaximumlikelihoodandmultiple imputationgivesaccurateestimationsunderMCARand MARmechanism.Asmissingdataarefrequently encounteredinbehavioral,psychologicalresearchmuch attentionhasbeengiventoanalyzestructuralequation models in the presence of missing data. Confirmatory factor analysis Themainimportanceinthedevelopmentanduseof measurement instruments is the degree to which they do measurethatwhichtheymeanttomeasure.Thatisto saysturcturesarevalid.Confirmatoryfactoranalysis (CFA)isamongthemostimportantmethodological approaches in order to analyze for the validity of factorial structures or within the framework of SEM (Byrne, 2001). CFAisusedforevaluatingandtestingthehypothesized factor structure ofscoresobtainedfromvarious Table1.Summaryofsamplesizesusedinmissingdata analysis. Percentage of missing data 0%1%5%10%20% 10099959080 200198190180160 500495475450400 1000990950900800 measurementinstrumentsandrelationsamonglatent constructs(forexample,attitudes,traits,intelligence, clinicaldisorders)incounselingandeducation(Sun, 2005;Jacksonetal.,2009).CFAgeneratesvarious statisticstoexplainhowwellthecompetingmodels explainedthecovariationamongthevariablesoffitthe data. Thesestatistics are called as fit statistics (Gillapsy, 1996). The correspondence between hypothesized latent variablemodelsisquantifiedbyfixindexes(Huand Bentler, 1995).Multipleimputation,rarelyusedforSEMs,isa method which is used for handling missing data but EM method is commonly used as missing data imputation and based on MLestimation.EMperformswellunderdifferentcondi-tionsinsimulationstudies(Zhang,2010).Regression imputation estimates missing values unbiased in the case of the data are MCAR (Tannenbaum, 2009). Determinationofrelationsamongvariablesorcon-structshasacrucialimportanceinthemeasurement approaches.Theseconstuctscanbeaffectedbymany threats.Threatstoreliabilitywillautomathicallyaffectto constructvalidity.Missingdataisoneofthethreatsto internal consistency reliability (Kenny, 2005). A measures internal consistency reliability is defined as the degree of ture score variation relative to observed-score variation: reliability of measure, ..true variability, ..observed variability, ..unexplained variability in the measure. Itcanbeseenfromtheequation,astheerrorvariance increases,reliabilitydecreases.Withrespecttomissing data,lostinformationcangiverisetolargeramountsof errorvariance.Missingdatahasnegativeeffectson research results such as contribution to biased resultsand makingitdifficulttomakevalidandefficientinferences aboutapopulation,decreasingstatisticalpowerand finallycauseviolationstatisticalassumptions(Schafer and Garham,2002;Kangetal.,2005;Kenny,2005;Kse211 Tannenbaum,2009;RoseandFraser,2008).Therehas beenconsiderableinterestintheeffectofthemissing datahandlingmethodsonstructuralequationmodeling (EndersandBandalos2001;Chenetal.,2012;Enders, 2001;Allison,2003).Consideringtheaforementioned research,missingdatahandlingmethodshaveeffecton goodnessoffitstatistics.Specifically,thisstudy specificallyaskswhichmissingdatahandlingmethod worksbestongoodnessoffitsttisticsinCFAwhen proportion of missing data and sample size are known? METHODOLOGY Data simulation The primary source of data used for statistical anlysis performed in thisstudywasasimulateddataset.R-studioprogramwasusedto generatedatasets.Reasonforusingsimulateddatawasthatitis difficulttosatisfyalloftheassumptionsunderexperimentalcon-ditionssuchasdifferentsamplesizesrangingfromverysmallto verylargewithdatamissingatdifferentratesinthesesamples.A secondreasonforusingsimulateddataisthatsincewestartwith completedataset,itisrelativelystraightforwardtoobservethe effectofmissingdataongoodnessoffitstatisticsbycomparing resultsdirectlybetweencompleteandincompletedatasets.This allowsonetoobjectivelyevaluatehowmuchoferrorcanbe corrected by using a particular missing data method.Four datasets (1 to 0)with 25 replications were simulated which included10continuousvariables(Table1).These10continuous variableshaveamultivatiatenormaldistribution.Foreaseof interpretation all variables were specified to have a mean of 0 and standard deviation of 1.Each ofthesesubsampleswasthenreduced insizeby1,5,10 and 20% in order to simulate datasets containing missing data. The caseswerediscardedrandomlyfromeachcompletesamples separetelyinordertomakesurethattherewerenodependencies betweensamples.Forexample,10caseswererandomlythrown outfromasamplesizen=100inordertoobtainapartialsample containing10%missingdata,n=90.Inordertoobtainasample with 20% missing data, 20 cases were randomly removed from the originalsampleofn=100againratherthanremoving10additional cases from the n=90 sample.

Method of analysis Missingdatahandlingmethods,listwisedeletion,regression imputation,EMimputation,andFIMLwereappliedtoallsamples containing missing data under CFA. The main consideration behind the choice of CFA was its widespread use among educational and psychological researchers. The data used in this study is simulated data.MCARpatterncanbeproducedbyrandomlydischarding cases. RESULTS Results of analytical procedures described in the methods sectionforthesimulateddataarepresentedinthis section. In order to see the effects of missing data hand-lingmethods,goodnessoffitindiceswerecalculated separately for each original data samples first (Table 2). Beforelookingat the relative performance ofvarious212 Educ. Res. Rev. Table 2. Goodness of fit indices for each sample sizes. Goodness of fit statistics Sample Size X2dfX2 /dfpRMSEAGFISRMRCFIIFIAGFI 10027,67320,860,6850,0000,950,0621,001,000,91 20032,42291,110,3020,0240,970,0470,980,980,94 50027,64350,790,8070,0000,990,0261,001,000,98 100037,59351,070,3510,0090,990,0221,001,000,99 missingdatahandlingmethods,subsamples(n=100, 200,500and1000)werethenreducedinsizeby1%, 5%,10%and20%inordertosimulatedatasetscon-taining missing data. The cases were discarded randomly from each complete sample for obtaining MCAR data. As aresult,16subsampleswithincomplatecaseswere obtained.Thesesubsampleswerethenhandledwith listwisemethodforeachpercentageofmissingand imputedwithmean,regression,andEMimputation methods.Forecahsamplesizeandpercentageof missingdata,datasetswereanaysedwithconfirmatory factor analysis and goodness of fit indices were presented with original samples in table 3, 4, 5, and 6.Whensamplesizeis100,forallpercentagesof missing, GFIs were presentedinTable3.Thefiguresin Table 3 show some important results. When proportion of missingdatais1%,listwisedeletionmethodhasshown visible performance.Namely, If the proportion of missing datais1%orsmaller,missingdatacasescanbe discardedfromthedataset.Because,thisproportionof missingnesshasnoeffectonGFIinconfirmatoryfactor analysis.Whenproportionofmissingdatais5%ormoreFIML methodshowsbeterfitindicesthanorthermethods. Besidesthis,listwisedeletionmethodworksworsethan orther methods beacuse CFA works well in large samples. Whensamplesizeis200,forallpercentagesof missing,GFIswerepresentedinTable4.Thevisible performance of EM and FIML imputation methods can be seeninallproportionsofmissingdata.Whenproportion ofmissingdatais1%,listwisedeletionmethodworks well.Also,itisworthwhiletopointherethatregression imputationmethodproducestheworstfitindices compared to other missing data handling methods. When sample size is 500, EM and FIML imputation methods are thebestmissingdatahandlingmethodsbecausethey producemoreacceptiblegoodnessoffitindicescom-pared to other missing data handling methods.Listwisedeletionmethodworkswellundercircum-stancesof1%percentageofmissingdata(Table5). Whensamplesizeis1000,forallpercentagesof missing, GFIs were presented in Table 6. The prominent performance of FIML imputation methos canbeseenin allproportionsofmissingdata.FIMLimputationmethod isthebestmissingdatahandlingmethodbecauseit producesmoreacceptiblegoodnessoffitindices.When proportionofmissingdatais1%,aswellasFIML imputationmethod,listwisedeletionmethodworkswell. Also,itisimportanttopointherethatregressionimpu-tation method generates the worst fit indices compared to other missing data handling methods. DISCUSSION AND CONCLUSIONS Theprimaryobjectiveofthisstudywastoexaminethe effect of missing data on goodness of fit statistics in SEM. For this aim, four missing data handling methods; listwise deletion, FIML, regression imputation and EM imputation wereexaminedwithsamplesizeandproportionof missingdata.Underthesmallsampleandlowmissing dataconditions,statisticalresultsimplythatlistwise deletionisoneofthesimplestandleastcomputation-intensive methods. Furthermore,listwisedeletionmethodisdefinitelynot recommendedforCFAanalysis,ifthesamplesizeis largeandmissingnessproportionishigh.Decreasing sample size by listwise deletion has negative effect on fit indices. Asthesamplesizeandproportionofmissingdata increases,FIMLimputationmethodsworksbestinall samplesizesandmissingdataproportions.Thisfinding was in line with Enders and Bandalos 2001.Regression impu-tationmethodswerenotgoodchoicesasmissing data handling method in the frame of results of this study. These results confirmed the findings of earlier researches ofZhang(2010)andTanenbaum(2009).Statisticalsoft-ware packages for SEM use maximum likelihood method andFIMLfortheestimationofmodelparameters.EM algorithmisageneralmethodfordoingMLandFIML estimationwithmissingdata.Simulationstudieshave shownthatittendstoperformwellundervarious conditions(Zhang,2010;EndersandBandalos(2001) 2001).Whenthemissingdatahandlingmethodis regressionimputationandthedataareMCAR,the resulting estimateswillberelativelyunbiasedinlarge Kse213 Table 3. GFIs for each percentage of missing (n=100) Sample Size MRMethodX2DfX2/df pRMSEASRMR CFIGFIAGFI100 1% Original27,67320,860,685 0,0000,0621,000,950,91 LWD27,76320,870,680 0,0000,0621,000,950,91 FIML28,89320,900,624 0,0000,0631,000,940,91 RI28,98320,910,624 0,0000,0631,000,940,91 EMI28,89320,900,624 0,0000,0631,000,940,91 5% Original27,67320,860,685 0,0000,0621,000,950,91 LWD26,81320,840,727 0,0000,0631,000,950,91 FIML27,92320,870,670 0,0000,0671,000,940,91 RI26,74320,840,729 0,0000,0621,000,950,91 EMI31,92321,000,470 0,0000,0670,990,940,90 10% Original27,67320,860,685 0,0000,0621,000,950,91 LWD23,84320,750,850 0,0000,0561,000,950,92 FIML26,68320,830,732 0,0000,0611,000,950,91 RI28,66320,900,636 0,0000,0621,000,950,91 EMI26,69320,830,732 0,0000,0611,000,950,91 20% Original27,67320,860,685 0,0000,0621,000,950,91 LWD25,16320,790,799 0,0000,0671,000,940,90 FIML27,86320,870,701 0,0000,0601,000,950,91 RI29,55320,920,590 0,0000,0631,000,940,90 EMI28,25320,880,656 0,0000,0631,000,950,91 Table 4. GFIs for each percentage of missing (n=200) Sample MRMethodX2dfX2/dfpRMSEASRMRCFIGFIAGFI Size 200 1% Original32,4229 1,120,3020,0240,0470,980,970,94 LWD33,0229 1,140,3000,0270,0480,970,970,94 FIML33,7529 1,160,2980,0280,0470,970,970,94 RI34,4129 1,190,2240,0310,0480,960,970,94 EMI35,7529 1,130,2880,0250,0470,970,970,94 5% Original32,4229 1,120,3020,0240,0470,980,970,94 LWD31,1129 1,070,360,020,0470,980,970,94 FIML33,1229 1,140,2990,0250,0480,970,970,94 RI33,9129 1,170,2420,0290,0480,970,970,94 EMI30,5629 1,050,3860,0160,0460,980,970,94 10% Original32,4229 1,120,3020,0240,0470,980,970,94 LWD36,6129 1,260,1560,0360,050,950,960,93 FIML33,7929 1,160,2790,0270,050,960,970,93 RI36,0729 1,240,1710,0350,050,960,970,93 EMI32,3329 1,110,2270,0270,0520,970,960,93 20% Original32,4229 1,120,3020,0240,0470,980,970,94 LWD34,9729 1,210,2050,0380,0480,970,970,94 FIML33,8329 1,160,2850,030,0490,960,970,94 RI37,6829 1,30,1290,0390,0520,940,960,93 EMI33,0529 1,140,2750,0290,0510,970,960,93 214 Educ. Res. Rev. Table 5. GFIs for each percentage of missing (n=500) Sample Size MRMethodX2dfX2/df pRMSEASRMRCFIGFIAGFI500 1% Original27,67 35 0,790,807 0,0000,0261,000,990,98 LWD28,00 35 0,800,566 0,0080,0221,000,990,99 FIML26,95 35 0,770,350 0,0090,0221,000,990,99 RI26,77 35 0,760,320 0,0100,0221,000,990,99 EMI26,92 35 0,770,366 0,0080,0221,000,990,99 5% Original27,67 32 0,860,807 0,0000,0261,000,990,98 LWD29,89 32 0,930,713 0,0000,0281,000,990,98 FIML26,89 32 0,840,835 0,0000,0261,000,990,98 RI31,39 32 0,980,643 0,0000,0291,000,990,98 EMI26,89 32 0,840,835 0,0000,0261,000,990,98 10% Original27,67 32 0,860,807 0,0000,0261,000,990,98 LWD31,43 32 0,980,641 0,0000,0291,000,990,98 FIML28,82 32 0,900,759 0,0000,0271,000,990,98 RI39,94 32 1,250,259 0,0170,0320,990,980,98 EMI28,88 32 0,900,758 0,0000,0271,000,990,98 20% Original27,67 32 0,860,807 0,0000,0261,000,990,98 LWD26,49 32 0,830,849 0,0000,0291,000,990,98 FIML28,69 32 0,890,821 0,0000,0271,000,990,98 RI37,20 32 1,160,367 0,0110,0310,990,990,98 EMI28,93 32 0,900,755 0,0000,0271,000,990,98 Table 6. GFIs for each percentage of missing (n=1000) Sample Size MRMethodX2dfX2/df pRMSEASRMRCFIGFIAGFI1000 1% Original37,5935 1,070,3510,0090,0201,000,990,99 LWD35,1835 1,050,4530,0080,0201,000,990,99 FIML32,1835 0,920,6040,0000,0201,000,990,99 RI51,0835 1,460,0380,0210,0300,990,990,98 EMI32,4235 0,920,5020,0070,0201,000,990,99 5% Original37,5935 1,070,3510,0090,0201,000,990,99 LWD41,8035 1,190,1990,0140,0200,990,990,99 FIML37,2035 1,060,3680,0080,0200,991,001,00 RI43,7935 1,250,1460,0160,0200,990,990,99 EMI38,1535 1,090,3800,0080,0201,000,990,99 10% Original37,5935 1,070,3510,0090,0201,000,990,99 LWD40,3235 1,150,2460,0130,0201,000,990,99 FIML38,9435 1,110,2960,0110,0201,000,990,99 RI50,6835 1,450,0420,0210,0301,000,990,98 EMI42,8335 1,220,1700,0150,0200,990,990,99 20% Original37,5935 1,070,3510,0090,0201,000,990,99 LWD55,0835 1,570,0480,0220,0300,990,990,98 FIML35,1835 1,050,3040,0000,0201,000,990,99 RI33,2635 0,950,5530,0000,0201,000,990,99 EMI34,4235 0,980,5020,0070,0201,000,990,99 samples though not fully efficient (Tannenbaum, 2009). The Americanpsychologicalassociation(APA)taskforce on statistical inference (1999) was against the appli-cation oftraditionalmethods,like listwise and pairwise deletion.Thesemethodswereamongthetheworst methodsavailableforpracticalapplications.Highlyre-commendedandpreferredmissingdatahandling methodsareEMimputationandmultipleimputation (SchaferandGraham,2002).Withtheframeofthe literatureandtheresultsofthisstudy,somerecommen-dations can be made; 1.Duringthedatacollection,researchersshouldbe careful to minimize missing values.2.Everyresearchershouldexaminethepatternsof missing values.3.Reasonsforandtheamountofmissingdatain research should be reported as well as how to handle missingness.4.FIMLimputationmethodshowsvisibleperformance and gives the closest fit indices to original fit indices. For this reason, FIML imputation method can be preferred in CFA. 5.Missingcasescanbeextractedfromthedatasetif missing data percentage is 1% or less. 6. This study can be replicated under larger sample sizes and proportions of missing data.7.Theeffectsofothermissingdatahandlingmethods, likemultipleimputationandfullinformationmaximum likelihoodongoodnessoffitindicesinCFAshouldbe examined. Therearesomelimitationsofthisstudy.Firstofall,this studywascomductedwithsimulateddataandmissing data pattern was assumed as MCAR. Other missing data patterns MAR and NMAR should be investigated. Conflict of Interests The author(s) have not declared any conflict of interests. REFERENCES AcockAC.(2005).Workingwithmissingvalues.J.MarriageFam. 67:1012-1028. AllisonPD(2003).Missingdatatechniquesforstructuralequation modeling. J. Abnormal Psychol. 112(4):545-557. ByrneBM(2001).StructuralequationmodelingwithAMOS,EQS,and LISREL: Comparative approaches to testing for the factorial validty of a measuring instrument.Int. J. Testing 1(1):55-86. CheemaJ(2012).Handlingmissingdataineducationalresearch usingSPSS.UnpublishedDoctoralDissertation.GeorgeMason University, USA.,ChenSF,WangS,ChenCY(2012).AsimulationstudyusinfEFAand CFAprogramsbasedontheimpactofmissingdataontest dimensionality. Expert Syst. Appl. 39:4026-4031. DaveyA,SavlaJ(2010).StatisticalpowerAnalysiswithMissingData.AStructuralEquationModelingApproach.Routledge.Taylor& Francis Group. New York, NY.EndersCK(2001).Aprimeronmaximumlikelihoodalgorithims availableforusewithmissingdata.StructuralEquationModeling:A Multidisciplinary J. 8(1):128-141. Kse215 EndersCK,BandalosDL(2001).Therelativeperformanceoffull informationMaximumlikelihoodestimationformissingdatain structural equation models. Struct. Equ. Modeling 8:430-457. EndersCK(2013).Dealingwithmissingdataindevelopmental research. Child Dev. Perspect. 7(1):27-31.GillapsyJA(1996).APrimeronConfirmatoryFactorAnalysis.Paper presented at the Annual Meeting of the Southwest Educational ResearchAssociation.NewOrleans,LA.(Ericdocument reproduction service no: ED 395 040). HuL,BentlerPM(1995).Evaluatingmodelfit.In:R.H.Hoyle(Ed.), StructuralEquationModeling:Concepts,IssuesandApplications Thousand Oaks, CA: Sage. pp.76-99. HuL,BentlerPM(1998)Fitindicesincovariancestructuremodeling: Sensivitytounparameterizedmodelmisspecification.Psychol. Methods 3:424-453Jackson DL, Gillapsy JA, Stephenson RP (2009). Reporting practices in Confirmatory factor analysis: An overview and some recomendations. Psychol. Methods 14 (1):6-23. JingL(2012).Missingdataimputation.UnpublishedDoctoral Dissertation. University of California. Los Angeles, USA. KangM,ZhuW,Tudor-LockeC,AinsworthB(2005).Experimental determinationofeffectivenessofanindividualinformation-centered approachinrecoveringstep-countmissingdata.Measurementin Physical Education and Exercise Sci. 9(4):233-250.KennyDA,McCoachDB(2003).Effectofthenumaberofvariableson measures of fit in structural equation modeling. Struct. Equ. Modeling 10 (3):333-351. KennyDA(2005).MissingData:MethodologyintheSocialSciences The Gulgord Press. New York, London. Mackelprang AJ (1970). Missing data analysis and multiple regression. Midwest J. Political Sci. 14 (3):493-505.MolenberghsG,Kenward MG (2007). Missing Data in Clinical Studies. John Willey & Sons, Ltd, UK. PeughJL,EndersCK(2004).Missingdataineducationalresearch:A reviewofreportingpracticesandsuggestionsforimprovement. Review of Educational Research, 74(4):525-556. PigottTD(2001).Arewievofmethodsformissingdata.Educational Research and Evaluation, 7, 353-383. Rose RA, Fraser MW (2008). A simplified framework for using multiple imputation in social work research. Soc. Work Res. 32(3):171-179. Saunders JA, Howell NM, Spitznagel E, Dork P, Proctor EK ,Pescarino R(2006).Imputingmissingdata:Acomparisonofmethodsforsocial work researchers. Soc. Work Res. 30(1):19-31Schafer J,Graham J (2002). Missing data: Our view on the state of the art. Psychol. Methods 7(2):147-177. SchlomerGL,BaumanS,CardNA(2010).Bestpracticesformissing datamanagementincounselingpsychology.J.Couns.Psychol. 57(1):1-10.Sterner WR (2011). What is missing in counseling research? Reporting missing data. J. Couns. Dev. 89:57-63.SunJ(2005).Assessinggoodnessoffitinconfirmatoryfactoranalysis. Meas. Eval. Couns. Dev. 37:240-256. Tabachnick BG, Fidell LS (2001). Using multivariate statistics (4th ed.). Needham Heights, MA: Allyn & Bacon. TannenbaumCE(2009).Theempiricalnatureandstatisticaltreatment ofmissingData.UniversityofPennsylvania,USA.Yaynlanmam Doktora Tezi. ZhangW(2010).Estimatinglatentvariableinteractionswithmissing data. Unpublished Doctoral Dissertation.Universityof Notre Dame. Indiana, USA.