You are on page 1of 12

SmallAreaEstimationTechniquefortheOklahoma

HealthCareInsuranceandAccessSurvey
November2009

Foreword
ThistechnicalreportdescribestheSmallAreaEstimationtechniqueemployedbySHADACforthe2008
OklahomaHealthCareInsuranceandAccessSurvey(OHIS).Theanalysiswasusedtodevelopcounty
levelestimatesofuninsurancebasedondatafromthe2008OHIS.Thisreportisadaptedfromthe
technicalappendixprovidedasadeliverablefortheproject.Notethatasimilaranalysiswasconducted
bySHADACforthe2004OHIS(reportedin2005)usingadifferentmethodology;acomparative
assessmentisincludedhere.

Introduction
Ingeneratingthese2008countylevelSmallAreaEstimation(SAE)estimatesoftherateofuninsurance
forthe77countiesinOklahoma,wehaveemployedamethodologicalapproachthatisbothnewand
significantlydifferentfromtheoneusedinour2005report.Thisreportdescribesthisnewapproachin
detailanddiscussesthereasonswhyweadoptedit.

Constraints
PertinenttotheapplicationofSAE,animportantfeatureoftheOHISinboth2004and2008hasbeenits
relativelysmallsurveysamplesize(3804nonelderlyin2008)relativetoitsnumerouscounties(77).For
example,in2008halfofthecountieshaveasamplesizeof25orfewer,30%have11orfewerand70%
havesamplesizesof47orfewernonelderlyrespondents.Smallsamplesizes(ornosampleatall)are,
ofcourse,anaturalcharacteristicofSAEapplications.These2008countyleveldataare,however,
sparse,andthathasimportantimplicationsfortheprecisionofanySAEestimatesaswellasfor
choosingthemethodologicalapproachthatsappropriateforderivingasetofpolicyusefulSAE
estimates.

EvaluationofSAEApproaches
Aspartofthis2009projecttogenerateSAEestimatesforthe77Oklahomacountiesweundertooka
rigoroussearchoftheliteratureonSAEmethodologies.Wealsoconductedanextensiveevaluationof
thenewcandidateforaSAEmethodologicalapproachthatweidentifiedfromtheliterature.We
conductedthisevaluationbyassessingtheSAEestimatesofthisnewmethodologyaloneandalso
relativetothesetofSAEestimatesthatweobtainedemployingtheoldmethodologyusedin2005.

NewSAEApproach
ThenewapproachidentifiedintheliteratureandthroughanInternetsearchisaBayesianSAEmodeling
approachthathasbeendevelopedandmadeavailablethroughaprojectcalledBIAS,shortforBayesian
methodsforcombiningmultipleIndividualandAggregatedataSourcesinobservationalstudies.This
projectisbasedattheDepartmentofEpidemiologyandPublicHealth,ImperialCollege,London.The
modelswehaveusedwerepresentedlastsummerasBayesianSmallAreaEstimationforpolicymaking
andpolicyassessment,byV.GomezRubio,N.Best,S.Richardson.andP.Clarke,alloftheDepartment
StateHealthAccessDataAssistanceCenter,November2009

page1

ofEpidemiologyandPublicHealth,ImperialCollegeLondon.Thispresentationwasmadeatthe
ResearchMethodsFestival,OxfordUniversity,England,3July2008.

TheBayesianstatisticalmethodologyhasbeenaroundforalongtime,since1763infact.Butithasonly
beensincetheadventofverypowerfulcomputersthatwereinexpensiveenoughtobewidelyavailable
thattheBayesianmethodologicalapproachhastakenoff.Indeed,itsburgeoningapplicationshave
broughtaboutamajorchangeinhowpeoplemakedecisionsinmanyareasanddisciplines,suchas
informatics,medicine,geneticsandtheInternet.Therearenowsoftwarepackagesdevotedexclusively
toBayesianmodeling,thebestknownofwhichisWinBUGS,whichstandsforWindowsbasedBayesian
InferenceUnderGibbsSampling.ItwasdevelopedbyateamofbiostatisticiansatCambridgeUniversity,
England.WeusedWinBUGStoestimateallourOklahomaSAEmodels.

TheunderlyingguidingprincipleofallSAEmodeling,andBayesianSAEinparticular,istheideaof
borrowingofpower,orborrowingofinformationfrommanysourcestoachievemorereliable
estimatesforsmallareaswithsmalldatarepresentationthanwouldbepossiblefromtheuseofjust
theseoftenverysmallsurveydatasamplesalone.InourBayesianmodelsthisborrowingofpower,or
borrowingofinformationtakesthreedistinctforms:

First,wehavesubstantiallymoreinformationaboutthetruevalueoftheuninsuranceratefor
theentirestatethanwehaveforindividualcountieswithinthestate,especiallysoforthe
smallestcounties.Bayesianmodelsoptimallybalancethereliabilityoftheseindividualdirect
estimatesofcountieswiththeirsometimesverysmallsamplesizeswiththemuchlarger
sampleavailableforthestateasawhole.ThusBayesianestimatesbythemselvesareanoptimal
blendoftheoftennotveryreliablecountyestimatesandthemuchmorereliablestatewide
estimate.InthiswaytheBayesiancountyspecificSAEestimatesofuninsuranceinvolve
borrowingofinformationfromthestatewiderate.Inparticular,whenthecountysurvey
sampleislarge,theBayesianSAEestimateforthatcountywillrelyheavilyonthedataforthat
largesamplecounty.Butwhenthecountysurveysampleisquitesmall,theBayesianSAE
estimateforthatcountywillshrinkthatsmallsamplecountyestimatetowardthestatewide
mean.Theamountbywhichthesmallsamplecountyestimateisshrunktowardthestatewide
meanisoptimallydeterminedbythesizeofthesample.

Asecondsourceofborrowedinformationcomesaboutfromtheestimationofregression
modelsthatincludevariablesthatpredictthelikelihoodofbeinguninsuredbasedontheentire
statesurveydataset.Thusthestrong,significantrelationshipbetweenthelikelihoodofbeing
uninsuredandpovertystatus,orincomelevelorothervariablescanbeusedtoborrow
informationfromtheserelationshipsandbringthemtobearonindividualcountiestoachieve
morereliableestimatesofSAEestimates.

FinallyandrestrictedtoBayesianSAEmodelsalonewehavemodelsthatnotonlyborrow
informationfromtheoverallstatewiderateofuninsuranceandfromtherelationshipbetween
predictorsofuninsuranceinourregressionmodels,butwecanalsoassesswhetherthepatterns
ofuninsuranceratesincountiessurroundingthecountyofinterestarestrongenoughtoallow
ustoborrowinformationfromthesegeographicallycloseareas.Thistypeofspatially
correlatedadjustmentstorateshasbeenveryeffectivelyappliedtotheestimationof
prevalenceratesforvariousdiseasesforsmallareas.Weusethismethodofspatially

StateHealthAccessDataAssistanceCenter,November2009

page2

correlatedadjustmentstoratesofuninsuranceinourBayesianmodelsaswell.Itis,again,
anothertypeofborrowedinformation.

EvaluationofAlternativeSAEModels
Forthesecondformofborrowedinformationthroughtheuseofaregressionmodelestimated
fromtheentiresurveydatasettherearetwobasictypesofmodelsthatonecanuse.Onecanusedata
onindividualrespondents(N=3804)andestimatetherelationshipatthatindividuallevelbetweenthe
probabilityofbeinguninsuredandvariouscharacteristicsmeasuredinthesurveysuchasemployment
status,educationlevelattained,andemployersize.[ThesearecalledUnitLevelmodels.]Oronecanuse
surveydataaggregatedtothecountylevel(N=77)fortheuninsuranceratesandusecountyleveldata
onpredictorsofcountyuninsurancerateslikecountyaverageincomeorcountyproportionofthe
populationbelowpoverty.[ThesearecalledAreaLevelmodels.]Ingeneral,eachtypeofmodelhasits
advantagesanddisadvantages.

However,intheparticularcaseofsurveydatasetsthathavemanyverysmallsamplesatthecounty
level(Area),AreaLevelmodelsprovideingeneralmorereliableestimates.Thisistruesinceitbecomes
ingeneralverydifficulttoreliablyestimatethemeansoftheexplanatoryvariableswithverysmall
samples.Forexample,theproportionofacountysresidentsemployedinlargeorsmallemployersif
estimatedwithverysmallnumbersofobservationscanandoftenwilldifferdramaticallyfromwhatthe
truecountymeanmightbeifalargersamplewereavailable.Wealsoobservedempiricallyinour
Oklahomadatathatthisdisadvantagewasveryimportantfortheactualcasesofmanysmallcounty
estimates.AreaLevelmodelsthatrelyondatafromoutsidethesurveyforexamplecensusdataon
povertyratesoraverageincomedonothavethisproblem.However,adisadvantagewiththeseArea
Levelmodelsthatuseexternaldataisthatingeneralyoucanonlyfindoneortwovariablesthatare
significantand,importantly,theexternaldataneedstocomefromayearrecentenoughtowhenthe
uninsurancedatawerecollectedtobeeffective.

Ourextensiveassessmentofthesetwotypesofmodelseasilysuggested,however,thattheAreaLevel
modelsprovidedbetterSAEestimates,andconsequentlyweusedanAreaLevelmodel.

AspartofourfullevaluationofSAEmodels,wealsoassessedtherelativeadvantagesofaBayesian
modelcomparedtotheSAEmethodologyusedinour2005Oklahomareport.Inthat2005SAEanalysis
weemployedaUnitLevelmodelwitharandomeffectforthecountyandweestimatedthismodelwith
thestatisticalpackageMLwiN.Usinganumberofcriteria,wejudgedthattheBayesianAreaLevelmodel
outperformedthepreviouslyemployedUnitLevelmodelwitharandomeffect(MLwiN).

InadditiontothisadvantageintermsofthegreateraprioriplausibilityofthecountySAEestimates
generatedbytheBayesianAreaLevelmodelcomparedtothepreviouslyemployedUnitLevelmodel
witharandomeffect(MLwiN)thereareotheradvantagesinusingaBayesianmodel.Specifically,with
aBayesianapproachonehasdirectmeasuresoftheuncertaintyofeachcountysSAEestimates.These
%torefertothe
BayesianmeasuresofuncertaintyarereferredtoasCredibleIntervals.1Ifweuse
unknown,trueuninsurancerateincountyA,thenwecandirectlydeterminethevalues
%

% forwhichwecansaythatthetrue,unknownvalueoftheuninsuredratein

BayesianCredibleIntervalsarenotthesamethingasconventionalConfidenceIntervals,buttheyfunctionin
approximatelythesameway.Theyallowformoremeaningfulstatementstobemadeaboutuncertaintythan
conventionalConfidenceIntervalsallow.
StateHealthAccessDataAssistanceCenter,November2009

page3

countyA,
%,hasa95%probabilityoffallingwithintheboundssetbythevalues
%

% .Moreformally,wecanidentifyfromtheBayesianresultsthevalues
%
. .

% forwhichthe
%
%
%

Asmeasuresofuncertainty,theseBayesianCredibleIntervalsintegrate,effectively,allsourcesofour
uncertaintyaboutourSAEestimates.Theyreflecttheuncertaintyabouttheunderlyingvaluesofthe
dataonthenumberuninsuredinthecounty.Theyalsoreflecttheuncertaintythatarisesconcerningthe
magnitudeoftheregressionmodelscoefficients,whichcanhaveimportantimpactsontheSAE
estimatesobtained.FinallyandinourspecialBayesianmodels,theyalsoreflecttheuncertaintyabout
thevaluesofthespatiallycorrelatedadjustmentsdiscussedabove.

Assuch,theseCredibleIntervalsprovideimportant,usefulguidesforpolicymakers.Itscriticalthat
policymakingbasedontheseSAEestimatesexplicitlyacknowledgethereliabilityoftheseSAEestimates
asexpressedinthevaluesoftheseBayesianCredibleIntervals.Putanotherway,thesizeofthese
BayesianCredibleIntervalsreflectthelimitstowhichconclusionscanandshouldbedrawnfromthese
SAEestimates.

Forseveralreasonsitwasnotpossibletoderiveusefulmeasuresofuncertaintyforthe2005SAE
estimatesinourearlierreport.Clearly,havingthecapabilitytoderivethesemeasuresforthepresent
studyfromourBayesianmodelsrepresentsanimportantimprovementinthepolicyutilityofourSAE
approach.

InadditiontotheseBayesianCredibleIntervals,theresanadditionalbenefitfromusingourBayesian
modelingstrategy.Specifically,asdescribedinmoredetailbelow,wehavedevelopedmodelsforboth
the2004dataandthe2008data.GivenourBayesianorientation,wehaveconstructedanoverallmodel
inwhichthe2004dataandthe2008dataareestimatedinparallel,whichyieldsSAEestimatesforboth
years.Giventheprovisionofthesenewestimatesfromthe2004dataandthe2008estimates,a
naturalquestionforpolicyiswhetherspecificcountieshaveexperiencedincreasesordecreasesintheir
SAEestimatesofuninsuranceoverthistimeperiod.Theadvantageofanoverallmodelwhichincludes
boththe2004and2008dataisthatthedifferenceincountyspecificestimatescanbedirectly
modeled.Consequently,weprovideestimatesoftheseestimateddifferencesfrom2004to2008in
countyspecificSAEestimatesofuninsurance.Inadditionandagaintoguidepolicymakersintheiruse
oftheseestimateddifferencesweprovideBayesianCredibleIntervalsprovidingouruncertaintyabout
thevaluesoftheseestimateddifferencesfrom2004to2008inSAEestimatesofuninsurance.Of
course,sincebothyearsofSAEestimatesarecombinedthroughthecreationofthisestimated
differenceofuninsurancerates,thedegreeofuncertaintyinthesedifferenceseffectivelycombinesthe
uncertaintyfromboth.Consequently,andalthoughmostoftheseBayesianCredibleIntervalsforthe
estimateddifferencesarelarge,theyprovideimportantstatementsofthelimitationsofwhatcanand
cannotbeconcludedfromthesecomparisons.

StateHealthAccessDataAssistanceCenter,November2009

page4

ModelDescription
In2005weusedtheentireOklahomasurveytogenerateourSAEcountyestimatesofuninsured,
includingthedatafromelderlyrespondents.SinceattherequestoftheStateweareestimatingour
SAEcountyestimatesofuninsuredin2008usingonlythenonelderlyandalsosincewehavechanged
SAEmodelingstrategies,wehavegeneratedsetsofcountySAEestimatesofuninsurancefromboththe
2008andthe2004surveys.

OurtwoSAEmodels,for2004and2008data,havethefollowingfeatures:

2005Model

Althoughweassessedanumberofvariablesfromcensusdataforinclusioninourmodel,only
theaverageincomeinthecountyandthe%ofthecountyspopulationbelowpovertywere
significantlyrelatedtothemeanuninsurancerate.Sincethesetwovariablesaresohighly
correlated,however,onlyoneofthemcouldbeincludedinourmodel.Afterevaluatingthesets
ofSAEestimatesgeneratedwithboth,wechosethemodelwithcountymeanincomein
thousands(IncomeK).Thetablebelowgivesthecoefficientvaluesandsignificanceforthis
model.

Coefficient
SD
RatioCoeff/SD
Intercept
0.12
0.35
0.339
IncomeK
0.0423
0.01
3.91

2009Model

Againweassessedanumberofvariablesfromcensusdataforinclusioninourmodel,butas
expectednoneweresignificantforthe2008countyuninsurancerates.Wefound,however,that
thecountyuninsuranceratein2004wasanimportantandsignificantpredictorofthecounty
uninsuranceratein2008.Consequentlyweusedthisvariabletoborrowinformationina
temporalsenseoverthis4yearperiod.Thetablebelowgivesthecoefficientvaluesand
significanceforthismodel.

Coefficient
SD
RatioCoeff/SD
Intercept
2.28
0.29
7.76
Unins_2004
3.47
1.28
2.71

StateHealthAccessDataAssistanceCenter,November2009

page5

SummaryStatisticsofResults
Whilewediscusstheresultsofbothsetsof2004and2008SAEestimatesindetailinthereport,herewe
presentseveralsummarymeasuresthatprovideimportant,relevantinformationwithwhichtoevaluate
theseBayesianSAEestimates.

WebeginwithdescriptivestatisticsfortheBayesianSAEestimatesofcountyuninsuranceratesin2008
andcomparethesewiththeestimatesgeneratedfromusingthe2008surveydatatodirectlyestimate
theseratesofuninsurance(referredtoasdirectestimates).[Theunitsoverwhichthesesummary
statisticsarecomputedarethecounties(N=77).]

SummaryStatisticsof2008DirectEstimatesandBayesian
SAEEstimatesofUninsuranceRates

DirectEstimate

BayesianSAEEstimates

Min

0.0%

12.3%

Max
Simple
Mean

80.0%

29.3%

19.1%

19.2%

Med

17.6%

18.8%

sd

15.0%

3.2%

Ascanbereadilyseen,thereisaverylargerangefromaminof0%toamaxof80%intheDirect
Estimatesofuninsuranceratesacrossthe77countiesin2008(oftenwithverysmallsurveysamples).
Thisis,ofcourse,tobeexpectedandisthereasonwhySAEmodelingtechniquesareundertaken.This
largevariabilityisalsoreflectedinthelargestandarddeviationforthesedirectestimatesof15%points.

Asisalsoreadilyseen,theBayesianSAEestimatesshrinkthisrangedowntoaminof12.3%andamax
of29.3%inaccordwiththeoptimalpropertiesofBayesianestimatesasdiscussedpreviously.Notice
thatthetwosimplemeans2ofthese77estimatesdonotdifferbetweenthedirectandtheBayesianSAE
estimates,asshouldbethecaseforanySAEestimator.

OnewayofevaluatingasetofSAEestimatesinvolvescomputingthefollowingteststatistic:
TakeeachcountysSAEestimateoftheuninsurancerateandmultiplythisbythecountyspopulation
andsumall77oftheseSAEestimatedcountynumberofuninsured.Thissumshouldcomeclosetothe
aggregatenumberofuninsuredderivedforthestateasawholebymultiplyingthesurveyweighted
estimate(18.9%)bythestatewidepopulation.ForourBayesianSAEestimates,thesetwoestimates
thecountypopulationweightedrateofSAEestimateduninsuranceandtheaggregatenumberof
uninsuredderivedforthestateasawholewereonly0.7%pointsdifferent.Thatis,adifferenceofless
thanonepercentagepoint.Incontrast,ourpreviouslyemployedUnitLevelmodelwitharandomeffect
(MLwiN)whenappliedtothe2008datayieldedacountypopulationweightedrateofuninsurance

Thesimplemeanofthese77countyDirectEstimaterates(19.1%)differssomewhatfromthesurveyestimate
(18.9%)byvirtueofthefactthatthesurveyestimateusesthesurveyweightsonall3804observations,and
consequentlyitisthecorrectestimatetouseforpublicreporting.Thatis,this19.1%rateisasimplemeanofthe
77countyrates,whichweprovidebecausewewanttoshowtheotherdescriptivestatisticsforthese77different
setsofestimates.Thisalsoappliestotheslightlyhigherrateof19.2%forthesimplemeanoftheBayesianSAE
estimates.
StateHealthAccessDataAssistanceCenter,November2009

page6

thatwasasmuchas13%pointslowerthanthesurveyweightedestimatetimesthestatewide
population.ThiswasonecriterionweusedtojudgebetweentheBayesianSAEmodelandour
previouslyemployedUnitLevelmodelwitharandomeffect(MLwiN).

WealsopresentsummarystatisticsforthewidthsoftheConfidenceIntervalsfortheDirectEstimates
andtheBayesianCredibilityIntervals.AnimportantfeatureofBayesianestimatesingeneralthat
appliesequallytoSAEestimatesisthattheynotonlyprovidemorereliablepointestimates,buttheir
levelofuncertaintyisusuallyloweraswell.Ascanbereadilyseen,thewidthsoftheCIsforthedirect
estimatesrangefromamaxof88%pointstoaminof6%points.TheBayesianCredibleIntervalshavea
maxwidthof24%pointsandalsoaminof6%points.Ofimportance,onaveragetheCIsforthedirect
estimateare2.5timeswider(37%vs.15%)thantheBayesianCredibleIntervals,asubstantialreduction
inuncertaintyfortheBayesianCredibleIntervals.

SummaryStatisticsofWidthsofConfidenceIntervals/CredibleIntervalsof
DirectEstimatesandBayesianSAEEstimatesofUninsuranceRates

DirectEstimateCIs

BayesianSAEEstimateCIs

Min

6%

6%

Max

88%

24%

Mean

37%

15%

Med

32%

15%

sd

19%

3%

Avisualsenseofwhatonecansayaboutsignificantdifferencesbetweencountiesintheir2008SAE
estimatesisprovidedbythesocalledcaterpillargraphbelow.Itgivesingraphicformatgoingfrom
thecountywiththelowestSAEestimateofuninsurancetothecountywiththehighestSAEestimateof
uninsuranceeachcountysSAEestimatetogetherwiththeCredibleIntervalbarforthatcountysSAE
estimate.Inthiscasetheverticallengthofthebarindicatesthelevelofuncertaintyforthisestimate.

Ascanbeseenfromthisgraphic,onecanmakestatementsconcerninghigherandlowerlevelsofSAE
uninsuranceratesforrelativelyfewcounties.Thatis,therearerelativelyfewpairsofcountiesforwhich
theCredibleIntervalbarsdonotoverlap.Thisisadirectconsequenceofthesparsedatadescribed
andthelevelofuncertaintythatnecessarilyinheresintheseestimatesofSAEuninsurancerateswhen
theyarederivedfromsmallcountylevelsurveysamples.

UsingtheseCredibleIntervalsweidentifywhichcountiesaresignificantlyhigherorlowerthanothers.

Astabulatedbelow,AdaircountysSAEUnins%rate,at29.3%,issignificantlyhigherthanthe7counties
listedinthe1sttable.InadditiontoCanadiancountyhavingoneofthose7SAEratesbelowAdair,
Canadian,at12.3%,isalsosignificantlylowerthan6additionalcounties,aslistedinthe2ndtable.

StateHealthAccessDataAssistanceCenter,November2009

page7

2008SAEUnin%&CI'sbyLowesttoHighest
45.0%
40.0%
35.0%
30.0%
High97.5%CI

25.0%

Low2.5%CI

20.0%

Unin%Rate

15.0%
10.0%
5.0%
0.0%

AdairCountysSAEUnins%
ExceedsthefollowingCounties

County
Adair

node
p[1]

Unins%
2008
29.3%

2.50%
21.3%

97.50%
39.5%

AdairhasasignificantlyHIGHERrateofUnins%
than:

Canadian

p[9]

12.3%

8.2%

16.8%

Woodward

p[77]

13.0%

7.0%

19.1%

Jackson

p[33]

13.6%

7.2%

20.0%

Cleveland

p[14]

15.6%

11.8%

19.8%

Tulsa

p[72]

16.4%

13.6%

19.6%

Grady

p[26]

14.7%

9.5%

20.1%

McClain

p[44]

14.7%

9.4%

20.5%

StateHealthAccessDataAssistanceCenter,November2009

page8

CanadianCountysSAEUnins%
IsBelowthefollowingadditionalCounties

County
Canadian

Node
p[9]

Unins%
2008
12.3%

2.50%

97.50%

8.2%

16.8%

CanadianhasasignificantlyLOWERrateof
Unins%than:
Cherokee

p[11]

25.3%

18.6%

32.8%

Okfuskee

p[54]

26.3%

18.2%

36.0%

Oklahoma

p[55]

21.8%

19.0%

24.8%

Pittsburg

p[61]

25.1%

18.0%

33.3%

Coal

p[15]

26.1%

17.0%

37.2%

Delaware

p[21]

24.3%

17.5%

32.7%

Forthe2004SAEestimates,thecaterpillargraphindicatessomewhatmoresignificantdifferentiation
amongthecounties.Inpartthisisduetothesomewhatlargersampleofnonelderly(N=4596)inthe
2004survey,approximately21%largerthanthe2008nonelderly.

2004SAEUnin%&CI'sbyLowesttoHighest
60.0%
50.0%

Unins%

40.0%
High97.5%CI
30.0%

Low2.5%CI
Unin%Rate

20.0%
10.0%
0.0%

Wesummarizethisgreaterdifferentiationinthetablebelowbygivingthenamesofthecountieswith
thelowestSAEUnin%rates,andforeachofthesenamedcountieswesimplyprovidethetotalnumber
ofcountiesthathadsignificantlyhigherSAEUnin%rates.
StateHealthAccessDataAssistanceCenter,November2009

page9


NumberofCountieswith2004SAEUnins%Rates
SignificantlyHigherthanthefollowingCounties

County

NumberofCounties
WithSignificantlyHIGHER
Unis%Rates

Canadian

27

Tulsa

16

Comanche

16

Cleveland

Rogers

Wagoner

Finally,wepresentthecaterpillargraphforthedifferencebetweenthe2008and2004countySAE
Unins%rates.Asnoted,ourBayesianmodelallowsustodirectlymodelthisdifference,and
consequentlyweprovideeachcountysdifferenceinSAEestimates(2008estimateminusthe2004
estimate)togetherwiththeCredibleIntervalbarforthatcountysdifferenceinSAEestimates.Again,
theverticallengthofthebarindicatesthelevelofuncertaintyforthisdifferenceinSAEestimates.Also
aswediscussed,sincebothyearsofSAEestimatesarecombinedthroughthecreationofthisestimated
differenceofuninsurancerates,thedegreeofuncertaintyinthesedifferenceseffectivelycombinesthe
uncertaintyfromboth.Consequently,mostoftheseBayesianCredibleIntervalsfortheestimated
differencesarelarge.Again,however,theyprovideimportantinformationforpolicymakerssincethey
indicatethelimitsofwhatcanandcannotbeconcludedfromthesecomparisons.

Inthiscase,eitheracountysdifferenceinSAEestimateshasaquitewideCredibleIntervalorwhen
narrower,thedifferenceitselfisquitesmall.Thus,all77countiesestimatesincludethezerovalueand
thusfornocountycanwesaythatitexperiencedeitherasignificantdecreaseorincreaseinitSAE
Unins%estimatebetweenthesetwosurveys.ThesummarystatisticsonthesedifferencesinSAE
estimatesarealsogiven.

StateHealthAccessDataAssistanceCenter,November2009

page10

2008 2004SAEUnins%EstimateDifference
0.200

2008 2004SAEUnis%

0.150
0.100
0.050
0.000

97.5%CI

0.050

2.5%CI

0.100

MeanDiff

0.150
0.200
0.250
0.300

DescriptiveStatisticsforCountyDifferencesin
SAEUnis%Estimates,20082004

Difference
Maximum
reduction
Maximum
increase

10.2%
3.4%

WenotethatifweweighteachcountysdifferenceinSAEestimatesbythecountypopulation,we
obtainaweightedmeandifferenceofa1.85%pointreduction.Thatis,the2008ratesareonaverage
1.85%pointslowerthanthe2004estimates.Ifwetakethedifferenceintheoverallsurveyweighted
meansoftheestimates,(18.9%20.7%),weobtainadifferenceofa1.87%pointreduction.Once
again,ourBayesianmodelspredictionsofthesecountydifferencesinSAEestimateswhenproperly
weightedbythecountypopulationyieldforallpracticalpurposesthesameresultsasthedifferences
inthetwosurveyweightedoverallmeans.Weprovidethecautionarynote,however,thatasimple,
unweightedmeanofthese77estimatesofdifferenceinSAEestimateswouldyieldanestimateof
averagechangethatislarger,butthisisbecauseitdoesnotusetheproperweightsandforthatreason
thissimplemeanshouldnotbetaken.Theindividualcountyestimatesofthisdifferenceareuseful,but
theirsimplemeanisnot.

SuggestedCitation
StateHealthAccessDataAssistanceCenter.2009.SmallAreaEstimationTechniquefortheOklahoma
HealthCareInsuranceandAccessSurvey.Minneapolis,MN:UniversityofMinnesota.

StateHealthAccessDataAssistanceCenter,November2009

page11

References
BayesianSmallAreaEstimationforpolicymakingandpolicyassessment,byV.GomezRubio,N.Best,
S.Richardson.andP.Clarke.Unpublishedpaperavailablefordownloadfromhttp://www.bias
project.org.uk/software/

WinBUGScodeforimplementingSAEmodelsisalsoavailablefordownloadathttp://www.bias
project.org.uk/software/

PowerPointslidesofthepresentationgivenbytheseresearchersattheResearchMethodsFestival,
OxfordUniversity,England3July2008,isavailableat
http://www.ncrm.ac.uk/RMF2008/festival/programme/spas/pres2/RMF08.pdf

Acknowledgement
ThisprojectwasfundedbytheOklahomaHealthCareAuthority(OHCA)andwasconductedunderthe
supervisionofKathleenThiedeCall,PhD,ofSHADAC.ThisreportwasauthoredbyGesturDavidson,
Ph.D.,ResearchAssociateatSHADAC.Dr.DavidsonwouldliketothankVirgilioGomezRubioforhelpful
adviceonimplementingtheirCARmodelusingWinBUGS.

Addendum
TheWinBUGScodeforthemodelusedforgeneratingthe2008SAEestimatesispresentedhere:

model
{
for( i in 1 : N ) {
r_dec[i] ~ dbin(p[i],n[i])

logit(p[i]) <- beta1 +beta2 * un2004bayes[i] +u[i] + v[i]


u[i]~dnorm(0, precu)
rankp[i]<-rank(p[1:N], i)
}
v[1:N] ~ car.normal(adj[], weights[], num[], precv)
precu~dgamma(0.01,0.01)
precv~dgamma(0.01,0.01)

beta1~dnorm(0,0.001)
beta2~dnorm(0,0.001)

StateHealthAccessDataAssistanceCenter,November2009

page12

You might also like