You are on page 1of 5

PEERGROUPMETHODOLOGIESANALYSISREPORT

SimingMao

PEERGROUPMETHODOLOGIES
CLUSTERINGALGORITHMS
Mostpeergroupmethodologiesarebasedonclusteranalysis.Clusteranalysisisanexploratorydataanalysis toolwhichaimsatsortingdifferentobjectsintogroupsinawaythatthedegreeofassociationbetweentwo objectsismaximaliftheybelongtothesamegroupandminimalotherwise.Giventheabove,clusteranalysis canbeusedtodiscoverstructuresindatawithoutprovidinganexplanation/interpretation.Inotherwords, clusteranalysissimplydiscoversstructuresindatawithoutexplainingwhytheyexist.Oneoftheadvantagesof clusteranalysisisthatwhengroupingunits,itcouldtakeseveralcharacteristicsofunitsintoconsideration. GeneralcategoriesofclusteranalysismethodsincludeJoining(TreeClustering),TwowayJoining(Block Clustering),andkMeansClustering.Anovelpeergroupmethodologyismentionedattheendofthissection. JOINING(TREECLUSTERING) Joining,alsoknownastreeclusteringisakindofhierarchicalclustering.Theobjectiveofthisalgorithmisto joinobjectstogetherintosuccessivelylargerclusters,usingsomemeasureofsimilarityordistance.The numberofclustersorthesizeofclustersmaydependontherequiredlevelofaggregation.Thismethod providesaneasywaytoseparatealargeclusterintotwoormoresmallerclusters.Oneweaknessofthis methodisitsrelativelylowertolerancetotherelationshipuncoveredatlaterstagesofanalysis. TWOWAYJOINING TwowayjoiningislesscommonlyusedcomparedtotreejoiningandkMeansClustering.Thismethodis usefulincircumstanceswhenitisexpectedthatbothcasesandvariableswillsimultaneouslycontributetothe uncoveringofmeaningfulpatternsofclusters.Whenwetrytoclusterhospitalsfrombothhospitalleveland departmentlevel,thismethodispromisingtogivearesultthatcombinesthesetwolevels,despiteofthe difficultywithinterpretingtheseresults.Theresultingstructure(clusters)isbynaturenothomogeneous. KMEANSCLUSTERING Thekmeansmethodwillproduceexactlykdifferentclustersofgreatestpossibledistinction.Thealgorithm willstartwithkrandomclusters,andthenmoveobjectsbetweenthoseclusterswiththegoalto1)minimize variabilitywithinclustersand2)maximizevariabilitybetweenclusters. Itshouldbementionedthatwiththismethodthebestnumberofclusterskleadingtothegreatestseparation (distance)isnotknownasaprioriandmustbecomputedfromthedata.Inotherwords,ifthereisa hypothesisorintuitionaboutnumberofclusters,kmeansmethodcouldprovideaquantitativewaytotest whetherthishypothesiswillhold.

STATISTICALSIGNIFICANCE Unlikemanyotherstatisticalprocedures,clusteranalysismethodsaremostlyusedwhenthereisnopriori hypothesis.Clusteranalysistriestofindthe"mostsignificantsolutionpossible."Therefore,statistical significancetestingisreallynotappropriatehere. ANOVELPEERGROUPMETHODOLOGY Similartotreeclustering,thenovelpeergroupmethodologypresentedherealsoclustersobjectsbasedon measureofsimilarityordistance.However,thedifferenceisthatthismethodresultsingroupswhereeach objectisatthecenterofitsownpeergroup.Possibleadvantagesoverotherpeergroupmethodologiesare thatfacilitiesareneverontheedgeofagroupandgroupsizeandthusgroupdispersionisdetermined bytheresearcher. Ontheotherhand,sinceapeergroupisformedbasedonwhichobjectisselectedasthecenter,thesetof peergroupswillbedifferentwhenanothersetofcenterobjectsareselected,whichmayleadtoinconsistent benchmarkingroupdiscussion. CASESTUDIES Threecasestudiesareprovidedheretoshowhowdifferentpeergroupmethodologiesareappliedinhealth caresettingsincludingagglomerativetreeclusteringmethod,kmeansmethod,andthenovelpeergroup methodmentionedabove.Therearealsosomediscussionsabouthowtoevaluatepeergroupsinthesecases.

CASE1:ANALTERNATIVEMETHODFORHOSPITALPARTITIONDETERMINATIONUSING HIERARCHICALCLUSTERANALYSIS
SAMPLES
150shorttermgeneralhospitals CHARACTERISTICS 1. WhetherornotthehospitalislocatedinanareadesignatedasaStandardMetropolitanStatistical Area(SMSA) 2. Whetherornotahospitalisaffiliatedwithamedicalschool 3. Thenumberoffacilitiesandservicesoffered 4. Themedianincomeinthehospitalscounty Quantify:Thesecharacteristicswereintendedtomeasurethosefactorswhicharebeyondahospitalscontrol butaffectcosts.

METHODOLOGY
ALGORITHM Treeclustering KEYTERMS Dendrogram:Thegrouphierarchyorthetreestructureresultedfromalgorithmsisreferredtoasa dendrogram. ExpectedDistinctiveness:Expecteddistinctivenessofapartitionisameasureusedtoevaluatetheefficiencyof aclusteringalgorithm.Theresultingdendrogramthatcouldmaximizetheexpecteddistinctivenessisdesired. STEPS 1. Analyzethefourmeasureswithprincipalcomponentsanalysis; 2. Computethesimilarityordistancebetweeneverytwohospitalswiththeresultingfactorscoresin previousanalysis; 3. Generatedendrogramswithclusteringmethodologies. 4. Findthedendrogramthatmaximizepartitionexpecteddistinctiveness. RESULT

CASE2:APPLICATIONOFCLUSTERANALYSISTOWARDSTHEDEVELOPMENTOFHEALTHREGION PEERGROUPS
SAMPLE
139healthregions CHARACTERISTICS 24variablesareusedtocoverasmanyoftheknownsocialandeconomicconstraintsandcriteriaplacedon theanalysis. Stepwisediscriminantanalysisoffinalhealthregiongroupingonall24variables

METHODOLOGY
ALGORITHM KMeansclustering STEPS 1. 2. 3. 4. 5.

Selectkobservationsasclusterseeds. Assignallobservationstotheirnearestclusterseed. Replaceclusterseedswiththemeanofallobservationsassignedtoeachcluster. Repeat1to3untilthechangeinclustermeansbecomesorapproaches0. Formfinalclustersbyassigningeachobservationtoitsnearestclusterusingthefinalclustermeans derivedfrom4.

RESULT Limitation:itisanaveragetechnique.Assuchmanyofthehealthregionswhichlieontheboundariesofa specificpeergroupcouldbepairedwithregionsinaneighboringcluster.Alsofromapracticalperspective,it wasrevealedthatmanyhealthregionsarenotthatdifferent.

CASE3:METHODTODEVELOPHEALTHCAREPEERGROUPSFORQUALITYANDFINANCIAL COMPARISONSACROSSHOSPITALS
SAMPLES
133VeteransAffairsmedicalcenters CHARACTERISTICS 1. Complexity(includingtertiarycare,referralcare) 2. Size(ofmedicalcenterandpatientpopulation) 3. Academicmission 4. Diseaseburdeninthepatientpopulation 5. PatientrelianceontheVAforhealthcare 6. Caredeliverystructure 7. Infrastructureatthemedicalcenterlevel 8. Communityandenvironment(densityandresourcesinthesurroundingcommunity)

METHODOLOGY
ALGORITHM Nontraditionalclusteranalysis,NearestNeighbors STEPS 1. 2. 3. 4.

Standardizeallvariables(includingbinaryones)tohaveameanof0andstandarddeviationof1. Calculatethedistancebetweenallhospitalsinmultidimensionalspace. Setcriteriafordeterminingthenumberofpeersinagroupandclusterthenearestneighbors. Inanotherway,selectaradiusindistanceandidentifyallmedicalcenterswithinthatdistanceorcloser aspeers.

KEYTERM: Squarerootofthesumofsquares(RSS):TheRSSofdistancesisametricthattakesintoaccountthedistances betweenallmembersofeachpeergroup,andisusedasanoverallmeasurethatrepresentsthediffusionof thepeergroup.Highervaluesindicateamorediffuseorlessdensegroup. RESULT

You might also like