Use of name recognition software, census data and multiple imputation to predict missing data on ethnicity: application to cancer registry records.

0.21 MB·English

by  Ryan, Ronan

Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3
http://www.biomedcentral.com/1472-6947/12/3
RESEARCH ARTICLE Open Access
Use of name recognition software, census data
and multiple imputation to predict missing data
on ethnicity: application to cancer registry
records
Ronan Ryan1*, Sally Vernon2, Gill Lawrence3 and Sue Wilson4
Abstract
Background: Information on ethnicity is commonly used by health services and researchers to plan services,
ensure equality of access, and for epidemiological studies. In common with other important demographic and
clinical data it is often incompletely recorded. This paper presents a method for imputing missing data on the
ethnicity of cancer patients, developed for a regional cancer registry in the UK.
Methods: Routine records from cancer screening services, name recognition software (Nam Pehchan and
Onomap), 2001 national Census data, and multiple imputation were used to predict the ethnicity of the 23% of
cases that were still missing following linkage with self-reported ethnicity from inpatient hospital records.
Results: The name recognition software were good predictors of ethnicity for South Asian cancer cases when
compared with data on ethnicity derived from hospital inpatient records, especially when combined (sensitivity
90.5%; specificity 99.9%; PPV 93.3%). Onomap was a poor predictor of ethnicity for other minority ethnic groups
(sensitivity 4.4% for Black cases and 0.0% for Chinese/Other ethnic groups). Area-based data derived from the
national Census was also a poor predictor non-White ethnicity (sensitivity: South Asian 7.4%; Black 2.3%; Chinese/
Other 0.0%; Mixed 0.0%).
Conclusions: Currently, neither method for assigning individuals to an ethnic group (name recognition and ethnic
distribution of area of residence) performs well across all ethnic groups. We recommend further development of
name recognition applications and the identification of additional methods for predicting ethnicity to improve
their precision and accuracy for comparisons of health outcomes. However, real improvements can only come
from better recording of ethnicity by health services.
Background the ethnicity of individuals; the use of Census data
This paper presents a method for imputing missing data based on area of residence to predict the ethnicity of
on the ethnicity of cancer patients, developed for a individuals; and finally, the use of multiple imputation
regional cancer registry in the UK. It implements exist- (MI) to make an allowance for the use of these predic-
ing approaches in a novel situation and evaluates their tors in subsequent statistical analyses. The method has
utility. It combines four differing approaches to dealing applications beyond cancer registries, and the results
with missing data of this type: the use of an additional presented below are relevant to all organisations and
source of self-reported ethnicity to replace the missing researchersthathaveincompleteinformationontheeth-
data; the use of name recognition software to predict nic group of individuals who have access to additional
datathatmayhelppredictethnicitywhenmissing.
TheWestMidlandsCancerIntelligenceUnit(WMCIU)
*Correspondence:[email protected] is a regional cancer registry covering a population of
1PublicHealth,EpidemiologyandBiostatistics,UniversityofBirmingham,
approximately5.3million.Theregistrycollectsinformation
BirminghamB152TTUK
Fulllistofauthorinformationisavailableattheendofthearticle
©2012Ryanetal;licenseeBioMedCentralLtd.ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommons
AttributionLicense(http://creativecommons.org/licenses/by/2.0),whichpermitsunrestricteduse,distribution,andreproductionin
anymedium,providedtheoriginalworkisproperlycited.
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page2of8
http://www.biomedcentral.com/1472-6947/12/3
about new cases of cancer and produces statistics about groups, or the original ethnic group of individuals who
incidence,prevalence,survivalandmortality:theavailabil- adopttheirpartner’ssurnamewherethepartnerisamem-
ityofinformationonthesociodemographiccharacteristics, berofanotherethnicgroup.Anotherapproachistheuse
includingethnicity,ofcancercasesisimportantforservice of census information on the ethnic distribution of the
planning, ensuringequalaccessto these services,and for area inwhichthe caselived.Examplesof thisare astudy
epidemiological studies. The mainsource of information ontheuptakeofbreastcancerscreeningservicesinLon-
ontheethnicityofcancercasesfortheWMCIUislinkage don,andthedevelopmentofariskcalculatorforcoronary
tothenationalHospitalEpisodeStatisticsdatabase(HES), heart disease based on data from a large set of general
whichroutinelycollectstheself-reportedethnicityofhos- practices [12,13]. This approach has the advantage of
pital patients [1]. However, HES currently only provides beingeasytoimplementifthepostcode oftheindividual
informationonpatientsadmittedtohospital,anddoesnot is available, but relies on the assumption that the ethnic
includepatientswhowereassessedortreatedinsecondary distributionofaCensusarea(approximately1,500people)
care,butnotadmitted,nordoesitincludeinformationon isanaccuratepredictoroftheethnicityofindividuals.
patientsseenprivatelyoutsidetheNationalHealthService The use of MI, such as that implemented by Royston
(NHS).Nationally,only80%ofcancercaseshaveethnicity andcolleaguesinthestatisticalsoftwarepackageStata,has
dataavailablefromHES[2].Thisrelianceonlinkagewith the potential to create a complete dataset that combines
hospitaladmissionshasimplicationsforcancerincidence, the predictions generated by the above methods, and
prevalenceandsurvival.Theuseofcompletecaseanalyses makes an allowance for the imprecision of these predic-
inthissituationhasthepotentialtocausebias:forexam- tionsthatiscarriedthroughtothefinalstatisticalanalyses
ple,casesofprostatecancerswhoarenottreated(’watchful [14].Itrequirestheusertohaveanaccurateunderstand-
waiting’)mayhavelongersurvivalthanthosewhosecancer ing of the reasons why the data are missing (the missing
ismore advancedand are therefore admitted to hospital. datamechanism),goodpredictorsofthevalueofthemiss-
The exclusion of the former from survival analyses will ingdata,andassumesthatthedataareotherwisemissing
tend to underestimate survival overall,and, ifethnicity is at random (MAR). It is, therefore, a combined approach
associated with treatment type, may obscure any differ- which aims to maximise accuracy where missing values
encesinsurvivalbetweenethnicgroups.Similarly,cancer arepredictedandtoadjusttheprecisionofanyestimates
caseswhoareknowntotheWMCIUthroughdeathcertifi- derivedfromthesepredictions(e.g.bywideningtheconfi-
cationonly(DCO) (theonlyinformationthey holdabout denceintervalsonestimatesofcancerincidence).
thecaseisbasedonadeathcertificate)havelesscomplete We will present information on the sensitivity and
linkagewithHES.Anycomparisonofcancerincidenceby specificity of each of these methods, and describe how
ethnicgrouprestrictedtocompletecaseshasthepotential these multiple sources were combined in a dataset that
toobscureanydifferencesbetweenethnicgroupsifpeople can be used for the estimation of cancer incidence and
from ethnic minorities are over represented among the survival by ethnic group.
DCOcases.
Therearealternativeapproachestodealingwithmissing Methods
informationthatallowallcasestoberetainedforanalysis. Population
The simplest of these is to link with additional external 111694cancercasesnormallyresidentintheWestMid-
information sources that record ethnicity, such as the landsregionoftheUKanddiagnosedbetween1January
NationalBreastCancerScreeningService(NBSS),butthe 2001and31December2007wereincludedinthecohort
populationcoverage,completenessandaccuracy ofthese forthisproject.Caseswerelimitedtothefivemostcom-
data may also be limited. A second commonly used mon cancer sites (breast, upper GI, lower GI, prostate,
approachisto uselists of namesthat areassociatedwith andlung)astherewouldbetoofewcasesfromthenon-
particular ethnic groups. Name recognition software Whiteethnicgrouptoallowreliablecomparisonsofinci-
packagessuchasNamPehchanandSANGRA havebeen denceandsurvivalforthelesscommoncancers.
usedinavarietyofsettingstoidentifypeoplewithSouth
Asian heritage [3-6]. More recently, another package Ethnic groups
called Onomap has become available: this attempts to The five ethnic groupings used were: White, South
identify people from many ethnic groups [7,8]. Each of Asian, Black, Chinese/Other, and Mixed. These broad
thesepackagesreliesonindividualshavinganamethatis groupings coincide with those used in official national
strongly associated with a particular ethnic group and statistics [15].
theirsensitivityandspecificityisknowntovaryfromset-
ting to setting [9-11]. None of them can identify people Availability of existing data on ethnicity
whowoulddescribethemselves as havinga mixedethni- Data on ethnicity were available for 85 506/111 694
city,individualswhosesurnamesarenotspecifictoethnic (77%) of cancer cases from HES, although some
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page3of8
http://www.biomedcentral.com/1472-6947/12/3
individuals had more than one ethnic group recorded. ordertoassesstheabilityofeachtoidentifytheethnicity
Where this occurred (1506/85 506 (1.8%)), we used the ofcancercases.
most commonly recorded ethnic group, in line with the
method recommended by Downing and colleagues [16]. Prediction of ethnicity using area-basedCensus data
This approach is believed to be appropriate as it uses Cases were assigned to the Census area (lower layer
most information. We set ethnic group to missing for super output area (LSOA): average size 1500 persons)
cases with more than one ethnic group recorded, but associated with their postcode or residence, and linked
without a ‘most common’ ethnic group (154/85 506 to a dataset with the ethnic distribution of each LSOA
(0.18%) of the complete cases), again in line with the in the 2001 national Census.
method used by Downing. We then tabulated the char-
acteristics of the cohort, and used a univariate chi- Full multiple imputation model
square analysis to identify demographic and clinical fac- Thelaststage inprocessing,theimputationofthe miss-
tors associated with missing ethnicity (the missing data ing ethnicity using the existing variables shown to be
mechanism) [17,18]. associated with missing ethnicity (Table 1) and external
information derived from the above sources was carried
Additional source of ethnicity data out in Stata using the MI package ICE [14]. The linked
The 28 795 breast cancer cases from the cohort were NBSS datawasuseddirectly to replacethe ethnicgroup
linked to data held by the eight breast cancer screening ofcasesnotalreadyknownfromlinkagewithHES,rather
services (NBSS) in the region using their NHS number. than as a separate predictor in the imputation model.
We assessed the value of this additional information by Where the case’s surname at birth was available (from
comparing sensitivity, specificity and positive predictive their death certificate) the Onomap and Nam Pehchan
value of the ethnicity recorded in the NBSS using the resultsforthatnamewereusedinplaceoftheresultsfor
HES dataset as the gold standard, where both were all known names in the imputation model, as name at
available. birth may be a more accurate reflection of ethnicity for
individualswhohavechangedtheirnamefollowingmar-
Prediction of ethnicity using name recognition software riage.Inaddition,asweintendedtousethedataforcan-
Twonamerecognitionapplicationswereavailableforuse cer-specific and all-cause mortality survival analyses, we
in this project: Nam Pehchan and Onomap. Nam Peh- included these survival outcomes as covariates in the
chan was used to identify people with South Asian imputationmodel, alongwith the time toeach outcome
names, and Onomap was used to identify people with (the Nelson-Aalen estimate of the cumulative hazard
names associated with White, South Asian, Black and function)[14].Thenumberofimputeddatasetswascho-
Chinese/Other ethnic groups. Asearly use ofNam Peh- sen conservatively: one imputed dataset per 1% of cases
chanwiththiscohortshowedthatitincludedforenames with any missing data. Missing ethnicity was imputed
which were common among other ethnic groups (e.g. using multinomial logistic regression within the ICE
‘Mona’),wedecidedtorunitonforenamesandsurnames package,andthedistributionofimputedvalueswastabu-
separately,ratherthanthedefaultapproachwhichwasto lated for comparison with the observed (complete case)
run it on all forenames and surnames combined. This data. The sensitivity, specificity, and positive predictive
allowedmatchedsurnamestocarryagreaterweightthan value (PPV) of the full multinomial logistic model was
matchedforenamesintheMIprocess. compared with the ethnic group recorded in HES in
Since Onomap only makes use of a single forename, ordertoassessitsability toidentifytheethnicity ofcan-
onlythefirstforenamewasusedwhentheapplicationwas cer cases. The model was developed on a randomly
run.However,ascasescouldhavemorethanonesurname selected 50% sample of the 85352 cases whose ethnicity
associatedwiththeirWMCIUrecord,theapplicationwas was recorded in the HES dataset. The remaining 50% of
runwitheachcombinationofforenameandsurnamefor cases were used to validate the model and derive the
each case. If Onomap assigned a case to more than one aboveestimates.The predictorsusedinthemodelwere:
broadethnicgroup,theirmultipleresultswerereplacedby ethnicity derived from name recognition software; Cen-
asingleresultaccordingtothefollowingorder ofprefer- sus estimates of ethnic distribution of population; num-
ence: Chinese/Other, Black, South Asian, White. This berofhospitaladmissions;yearofdiagnosis;patientseen
orderingcorrespondswiththerelativesizeofeachgroup outside the NHS (yes/no); screen-detected cancer (yes/
within the regional population, with preference given to no); death certificate only cancer registration (yes/no);
thelesscommonethnicgroups.Thesensitivity,specificity, cancertreatmenttype(surgery/radiotherapy/chemother-
andpositivepredictivevalue(PPV)ofthetwoapplications apy); deprivation score; gender; age at diagnosis; cancer
wascompared withthe ethnic group recorded inHES in site; and death during follow-up period (all-cause and
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page4of8
http://www.biomedcentral.com/1472-6947/12/3
Table 1Characteristics ofthe ProjectCohort
Cases Caseswithmissingethnicityfollowing Chi2statistic
linkagewithHESrecords(%ofcases) (P-value)
Site LowerGI 24446 4112(17)
Breast 28795 6029(21)
Lung 24060 5660(24)
Prostate 23716 8814(37)
UpperGI 10677 1727(16) 3500(p<0.001)
Yearofdiagnosis 2001 15102 4118(27)
2002 15523 3840(25)
2003 15731 3772(24)
2004 16162 3681(23)
2005 16317 3632(22)
2006 16458 3470(21)
2007 16401 3829(23) 206(p<0.001)
Deprivation 1(mostdeprived) 26738 5149(19)
(IncomeDomainof 2 22104 4856(22)
IndexofMultiple 3 22759 5425(24)
Deprivation2007) 4 22465 5942(26)
5(leastdeprived) 17628 4970(28) 6300(p<0.001)
Age <40 1771 251(14)
40-49 5622 892(16)
50-59 15338 2933(19)
60-69 27759 5924(21)
70-79 35258 8538(24)
80+ 25946 7804(30) 1100(p<0.001)
Sex Male 59592 15454(26)
Female 52102 10888(21) 391(p<0.001)
DeathCertificate No 106217 23577(22)
Onlyregistration Yes 5477 2765(50) 2300(p<0.001)
Everseenprivately No 106566 23113(22)
(cancerwasdiagnosedortreatedoutsidethefree Yes 5128 3229(63) 4600(p<0.001)
NationalHealthServiceatleastononeoccasion)
Surgery No 58875 18344(31)
Yes 52819 7998(15) 4000(p<0.001)
Radiotherapy No 73520 19009(26)
Yes 38174 7333(19) 616(p<0.001)
Chemotherapy No 93778 24650(26)
Yes 17916 1692(9) 2400(p<0.001)
Screendetected No 22900 5012(22)
breastcancer* Yes 5895 1017(17) 61(p<0.001)
HES-linked No 19694 19694(100)
Yes 92000 6648(7)
Numberofadmissions 0 19694 19694(100)
(includesnon-canceradmissions) 1 8012 2071(26)
2 10261 1414(14)
3 10523 985(9)
4 9332 562(6)
5+ 53872 1616(3) 6300(p<0.001)**
*Comparisonofbreastcancercaseswhowereandwerenotdetectedbypopulationscreening.
**Chi-squaretestexcludescaseswithnoadmissionsas,bydefinition,nonehaveethnicityrecorded.
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page5of8
http://www.biomedcentral.com/1472-6947/12/3
duetoprimarycancerseparately)andtimetodeath/cen- The sensitivity of Onomap is high for White and South
soring(Nelson-Aalencumulativehazard). Asianethnicgroups(99.8%and82.1%),butlowforBlack
andChinese/Othergroups(4.4%and0.0%).Thesensitiv-
Research governance ityofNamPehchanwaslowerthatofOnomapforSouth
The project did not require separate ethical approval as Asiancases(71.1%and82.1%),butwhenbothwerecom-
it was commissioned by, and carried out in collaboration bined,sensitivitywashigherthaneachindividualapplica-
with, the regional cancer registry. Cancer registries have tion (90.5%). A total of 14 615 cases had their name at
legal support to collect data relating to cancer under birthrecordedontheirdeathcertificate.
Section 251 of the NHS Act 2006 (and formerly under Table 5 shows the sensitivity, specificity and PPV of
Section 60 of the Health and Social Care Act 2001). 2001 national Census data on ethnicity as a predictor of
[http://www.ukcancassoc.ismysite.co.uk/content/legal- the ethnic group of individual cases. The sensitivity of
background#S251]. Census data for the White ethnic group is high (99.3%),
butverylowforallotherethnicgroups(lessthan7.4%for
Results South Asian cases, 2.3% for Black cases, and 0% for the
The completeness of information on the ethnicity of remainingtwogroups).
cancer cases following linkage with HES varied signifi- Theethnicityofcasesthatweremissingfollowinglink-
cantly (P < 0.001 in each case) by the demographic and age with the HES and NBSS datasets was imputed in
clinical factors listed in Table 1. Stata usingICE withanimputationmodel thatincluded
The value of linkage with breast cancer screening ser- the variables significantly associated with missingness
vices(NBSS)informationonethnicityisshowninTable2 (Table1),thepredictedethnicityofeachcasemadeusing
and Table 3. Table 2 describes the sensitivity, specificity Onomap and Nam Pehchan, and the ethnic breakdown
and PPV of ethnicity derived from the NBSS compared of the area of residence of the case. The number of
withthatrecordedinHES(i.e. usingHESasa goldstan- imputeddatasetsgeneratedforthefullrunwassetto23
dard), for 5243 breast cancer cases with ethnic group asethnicitywasmissingfor22.6%ofthecases(Table3).
recordedinbothHESandNBSS datasets.Sensitivitywas Table6showsthesensitivity,specificityandPPVofthe
high(>90%)forWhiteandSouthAsiancases,andmod- fullmultinomiallogisticregressionmodelusedtoimpute
erately high (61.4%) for Black cases, suggesting that the missingethnicity.Thesensitivityandspecificityofthefull
NBSScouldbeused to determinethe ethnicityof cancer modelwascomparabletothatfromthenamerecognition
registry cases where this was not recorded in HES. No software alone for the White group (99.3%/56.0% vs.
cases recorded as Chinese/Other or Mixed ethnicity in 99.8%/51.5%, respectively). The sensitivity of the full
HES were assigned the same ethnic group in the NBSS modelwasslightlyhigherforcasesfromtheSouthAsian
dataset.However,thevalueoftheNBSSrecordsforthese group than name recognition software alone (94.7% vs.
twoethnicgroupscannotbepreciselydeterminedbecause 90.5%,respectively),andsubstantiallyhigherforBlackand
of the small numbers involved (14 and 10 cases, respec- Chinese/Otherethnicgroups(20.4%vs.2.3%and21%vs.
tively).Table3showstheeffectofusingtheNBSSdatato 0%, respectively).The sensitivityofthefullmodel forthe
resolve the ethnicity of registry cases that were not Mixedethnicgroupremainedat0%.
recordedinHES.Atotalof1082/26342(4.1%)breastcan- Table7comparestheproportionofcasesineachethnic
cer cases whose ethnicity was not known in HES had an groupforcompleteandimputedcases(all23imputations
ethnic group recorded in the NBSS. Overall it decreased combined). The proportion of cases in the White, South
the proportion of cancer cases with unknown ethnicity Asian and Black groups was slightly lower among the
from23.6%to22.6%. imputed cases than the complete cases (95.8% vs. 96%,
The sensitivity, specificity and PPV of Onomap and 1.7% vs. 1.8%, and 1.6% vs. 1.7%, respectively). For the
NamPehchanforeachethnicgroupisshowninTable4. remaining ethnic groups, the proportion of cases each
Table 2Sensitivity, Specificity and Positive Predictive Value ofNBSS-derived Ethnicity forBreast Cancer Cases
EthnicgrouprecordedinHES NumberofcasesrecordedinHES NBSS
Sensitivity Specificity Positivepredictivevalue
White 5093 99.7% 77.3% 99.3%
SouthAsian 82 90.0% 99.8% 87.1%
Black 44 61.4% 99.9% 79.4%
Chinese/Other 14 0.0% 100.0%
Mixed 10 0.0% 100.0%
Includes5243caseswhereethnicgroupwasrecordedinbothHESandNBSSdatasets.Individuallogisticmodels(positiveoutcomethreshold:p>=0.5).
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page6of8
http://www.biomedcentral.com/1472-6947/12/3
Table 3Ethnicity ofCases FollowingLinkage withHES andNBSSDatasets
Ethnicgroup Caseswithethnicgroup CaseswithunknownethnicityresolvedbyNBSSlinkage Ethnicbreakdownofcohort
recordedinHES(%) followinguseofHESandNBSS(%)
White 81934 (73.4) 1053 82987 (74.3)
SouthAsian 1545 (1.4) 13 1558 (1.4)
Black 1429 (1.3) 11 1440 (1.3)
Chinese/Other 303 (0.3) 3 306 (0.3)
Mixed 141 (0.1) 2 143 (0.1)
Notknown 26342 (23.6) 26331 (22.6)
Total 111694 1082
group was approximately 1.5 times higher among the ethnicityratherthanincludingitasapredictorofethni-
imputed cases (0.6% vs. 0.4%, and 0.3% vs. 0.2% for Chi- city in the MIprocess: it didrefer directly to the person
nese/OtherandMixedethnicgroups,respectively.) of interest. Similar datasets were not available for the
othercancersitesofinterest.
Discussion TheperformanceofNamPehchaniswidelyknownbut,
The main aim of this project was to create a method to asfarasweareaware,thisisthefirstpeerreviewedreport
imputetheethnicgroupofcancercaseswhowerenotified on the Onomap application. The higher sensitivity and
to the regional cancer registry, but whose ethnic group specificity that was achieved by using both applications
wasnotavailablefromtheirmainsource,linkagewiththe togethersuggeststhatthebestname-basedpredictionsof
nationaldatabaseonhospitaladmissions(HES).Wemade ethnicity can be achieved by the use of multiple applica-
use of precise external information on the ethnicity of tions.It is, however,unlikely that namerecognition soft-
caseswherepossible,throughlinkagewithafurtherdata- ware will ever precisely predict membership of some
set (the NBSS), two name recognition applications, and ethnic groups: although many people from South Asian,
area-basedinformationontheethnicmake-upoftheresi- Chineseandsomeother ethnicgroupsmayhavedistinc-
dent population. We then assessed the value of each of tive names, many individuals from White, Black and
theseadditionalsourcesbycomparingthemwiththeeth- Mixed ethnic groups do not. This suggests that we will
nicgroupofcaseswhoseethnicitywasknownfromHES. alwayshavetomakesomeallowancefortheirimprecision,
In the final stage of the method, we created a dataset andincludeadditionalpredictorsofethnicity.
whichcanbeusedtoestimateethnicgroupspecificcancer Area-based information from the national Census is a
incidenceandsurvival:thisinvolvedtheuseofaMIpro- popular predictor of ethnicity and easy to implement,
cedure(ICE). but this project demonstrates that is not precise enough
The main benefit of using additional linked datasets, to be used alone. Although sensitivity for the White eth-
liketheNBSS,isthatitmakesuseofpreciseinformation nic group may be high (> 99%) specificity is very low
recordedabouttheindividualsofinterest.Themainlim- (21%), showing that it misclassifies approximately 4 out
itationforthisprojectisthattheNBSSdatasetonlycon- of every 5 people from non-White backgrounds as
tains breast cancer cases who attended the screening White. Similarly, sensitivity is poor for the most com-
programme andwho had theirethnic grouprecordedat mon non-White ethnic groups in the region: the use of
thattime:thisresolvedtheethnicgroupofjust1%ofthe area-based Census data only appears to correctly iden-
cancer cases whose ethnicity was not already known tify approximately 7% of South Asian and 2% of Black
from linkage with HES. We decided to use NBSS cases. Setting alternative cutoff values for the model pre-
recorded ethnicity as a direct substitute for missing dictions from the default value of 0.5 did not improve
Table 4Sensitivity, Specificity and Positive Predictive Value ofNameRecognition Software
Namerecognitionsoftware Ethnicgroup Sensitivity Specificity Positivepredictivevalue
Onomap White 99.8% 51.5% 98.0%
SouthAsian 82.1% 99.9% 92.9%
Black 4.4% 99.9% 70.8%
Chinese/Other 0.0% 100.0%
NamPehchan* SouthAsian 71.1% 99.9% 94.5%
OnomapandNamPehchancombined SouthAsian 90.5% 99.9% 93.3%
Includes85352caseswhereasingleethnicgroupwasrecordedintheHESdataset.Individuallogisticmodels(positiveoutcomethreshold:p>=0.5).
*Matchedonforenameandsurnameseparately.
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page7of8
http://www.biomedcentral.com/1472-6947/12/3
Table 5Sensitivity, Specificity and Positive Predictive Table 7Comparison ofDistribution ofEthnic Groups:
Value ofCensus Data onEthnicity Observedand Imputed
Ethnicgroup Sensitivity Specificity Positivepredictivevalue Ethnicgroup Observed% Imputed*% Total*%
White 99.3% 21.4% 96.8% White 96.0 95.8 96.0
SouthAsian 7.4% 99.8% 44.9% SouthAsian 1.8 1.7 1.8
Black 2.3% 99.9% 34.4% Black 1.7 1.6 1.7
Chinese/Other 0.0% 100.0% Chinese/Other 0.4 0.6 0.4
Mixed 0.0% 100.0% Mixed 0.2 0.3 0.2
Includes85352caseswhereasingleethnicgroupwasrecordedintheHES Total 100 100 100
dataset.Censusdatausedtopredictethnicgroupwere:percentageoflocal
*Usingall23imputationscombined.
populationinSouthAsian,Black,Chinese/OtherandMixedethnicgroupat
lastnationalcensus.Individuallogisticmodels(positiveoutcomethreshold:p
>=0.5). identifymembershipofthesegroupsmoreprecisely,using
age-specificdata ontheethniccompositionofsmallgeo-
the predictive performance of the Census data to any graphic areas (LSOAs) if this is published following the
great extent (results not shown). nextnationalCensusorcountryofbirth,forexample.
The full model used to predict ethnicity within the MI
procedure did appear to be superior to naming software Conclusions
and Census data alone. The model appeared to perform In summary, we have developed a method and dataset
best for the South Asian ethnic group, and did identify that will allow comparison of cancer incidence and sur-
membershipoftheWhite,BlackandChinese/Othereth- vival between ethnic groups. However, the sensitivity of
nic groups with greater sensitivity than any of its indivi- the Onomap name recognition application appears to be
dualconstituentpredictors.However,thesensitivityofthe low for people from non-White and non-South Asian
fullmodelinabsolutetermsforBlack,Chinese/Otherand ethnic groups, suggesting that it is of limited use for
Mixed groups is low. It is, therefore, uncertain that the studies that wish to classify individuals by ethnic group.
existingpredictors can be improved,or that new predic- Similarly, area-based information from the national Cen-
torscouldbeadded, whichwouldincreasethesensitivity sus, a common approach where individual names are
of future models to levels similar to that seen for the not available but area of residence is, appears to be
SouthAsiangroup. imprecise, particularly for the less common ethnic
The greatest difference between the observed and groups. Currently, neither method for assigning indivi-
imputed data in the final model was for two ofthe three duals to an ethnic group performs well across all ethnic
minority groups whose ethnicity is poorly predicted by groups. We recommend further development of name
name recognition software (Chinese/Other and Mixed). recognition applications and the identification of addi-
This may indicate that the imputation process does not tional methods for predicting ethnicity to improve their
performwellinthesecases.New MI modelsmay benefit precision and accuracy for comparisons of health out-
from including other predictors of ethnicity which help comes. Nevertheless, neither imputation nor name
recognition software will be completely accurate: reliable
statistics relating to the incidence, prevalence and survi-
Table 6Sensitivity, Specificity and Positive Predictive
val of persons with cancer by ethnic group require more
Value ofFull Model
complete recording of these data [19].
Ethnicgroup Sensitivity Specificity Positivepredictivevalue
White 99.7% 56.0% 98.2%
Acknowledgements
SouthAsian 94.7% 99.8% 90.4%
ThisprojectwassupportedbytheWestMidlandsCancerIntelligenceUnit
Black 20.4% 99.8% 63.6% (WMCIU).ItwascommissionedbytheWMCIUwhichprovidedaccesstothe
Chinese/Other 21.0% 99.9% 57.6% dataandfundingforRRtocarryouttheproject.GListheDirectorofthe
WMCIUandSVwastheDeputyDirectorofCancerRegistrationatthe
Mixed 0% 100%
WMCIUatthetimeofthestudy:theycontributedtothewritingofthe
Amultinomiallogisticregressionmodelwasusedtopredictethnicgroup.The manuscriptandtothedecisiontosubmitthemanuscriptforpublication.RR
modelwasdevelopedonarandomlyselected50%sampleofthe85352cases andSWareemployedbytheUniversityofBirmingham.
whoseethnicitywasrecordedintheHESdataset.Theremaiming50%of
caseswereusedtovalidatethemodelandderivetheaboveestimates.The Authordetails
predictorsusedinthemodelwere:ethnicityderivedfromnamerecognition 1PublicHealth,EpidemiologyandBiostatistics,UniversityofBirmingham,
software;Censusestimatesofethnicdistributionofpopulation;numberof BirminghamB152TTUK.2EasternCancerRegistrationandInformation
hospitaladmissions;yearofdiagnosis;patientseenoutsidetheNHS(yes/no);
Centre,PublicHealthBuilding,UnitC-MagogCourt,ShelfordBottom,
screen-detectedcancer(yes/no);deathcertificateonlycancerregistration HintonWay,Cambridge,CB223ADUK.3WestMidlandsCancerIntelligence
(yes/no);cancertreatmenttype(surgery/radiotherapy/chemotherapy);
Unit,PublicHealthBuilding,UniversityofBirmingham,BirminghamB152TT
deprivationscore;gender;ageatdiagnosis;cancersite;anddeathduring
UK.4PrimaryCareClinicalSciences,UniversityofBirmingham,Birmingham
follow-upperiod(all-causeandduetoprimarycancerseparately)andtimeto
B152TTUK.
death/censoring(Nelson-Aalencumulativehazard).
Ryanetal.BMCMedicalInformaticsandDecisionMaking2012,12:3 Page8of8
http://www.biomedcentral.com/1472-6947/12/3
Authors’contributions 19. IqbalG,GumberA,SzczepuraA,JohnsonM,WilsonS,DunnJ:Improving
Allauthorscontributedtothedesignoftheproject.RRledthe ethnicitydatacollectionforcancerstatisticsintheUK.DiversityinHealth
developmentofthemethodforimputingmissingdata,performedthe andCare2009,16:267-285.
analysis,andwrotethefirstdraftofthemanuscript.Allauthorscontributed
tolaterdraftsandapprovedthefinalmanuscript. Pre-publicationhistory
Thepre-publicationhistoryforthispapercanbeaccessedhere:
Competinginterests http://www.biomedcentral.com/1472-6947/12/3/prepub
Theauthorsdeclarethattheyhavenocompetinginterests.
doi:10.1186/1472-6947-12-3
Received:18May2011 Accepted:23January2012 Citethisarticleas:Ryanetal.:Useofnamerecognitionsoftware,
Published:23January2012 censusdataandmultipleimputationtopredictmissingdataon
ethnicity:applicationtocancerregistryrecords.BMCMedicalInformatics
andDecisionMaking201212:3.
References
1. HospitalEpisodesStatistics.[http://www.hesonline.nhs.uk],accessed01
Sept2010.
2. NationalCancerIntelligenceNetworkCoordinatingCentre:Cancerincidence
andsurvivalbymajorethnicgroup,England,2002-2006[http://publications.
cancerresearchuk.org/downloads/Product/CSINCSURVBYETHNICITY.pdf],
accessed13Jan2012.
3. CumminsC,WinterH,ChengKK,MaricR,SilcocksP,VargheseC:An
assessmentoftheNamPehchancomputerprogramforthe
identificationofnamesofsouthAsianethnicorigin.JPublicHealthMed
1999,21:401-6.
4. PriceCL,SzczepuraAK,GumberAK,PatnickJP:Comparisonofbreastand
bowelcancerscreeninguptakepatternsinacommoncohortofSouth
AsianwomeninEngland.BMCHealthServicesResearch2010,10:103.
5. SzczepuraA,PriceCL,GumberA:Breastandbowelcancerscreening
uptakepatternsover15yearsforUKSouthAsianethnicminority
populations,correctedfordifferencesinsocio-demographic
characteristics.BMCPublicHealth2008,8:1471-2458.
6. NanchahalK,MangtaniP,AlstonM,dosSantosSilvaI:Developmentand
validationofacomputerizedSouthAsianNamesandGroup
RecognitionAlgorithm(SANGRA)foruseinBritishhealth-relatedstudies.
JPublicHealthMed2001,23:278-85.
7. OnoMAP.[http://www.onomap.org],accessed5Oct2010.
8. TheNamesProjects.[http://redress.lancs.ac.uk/resources/launch.php?
creator=MateosPablo&title=TheNamesProjects],accessed13Jan2012.
9. MateosP:Areviewofname-basedethnicityclassificationmethodsand
theirpotentialinpopulationstudies.PopulationSpaceandPlace2007,
13:243-263.
10. BrantLJ,BoxallE:Theproblemwithusingcomputerprogrammesto
assignethnicity:Immigrationdecreasessensitivity.PublicHealth2009,
123:316-320.
11. IdentifyingEthnicity:comparisonoftwocomputerprogrammes.[http://
www2.warwick.ac.uk/fac/med/research/csri/ethnicityhealth/
aspects_diversity/identifying_ethnicity/identifying_ethnicity.doc].
12. RenshawC,JackRH,DixonS,MøllerH,DaviesEA:Estimatingattendance
forbreastcancerscreeninginethnicgroupsinLondon.BMCPublic
Health2010,10:157.
13. Hippisley-CoxJ,CouplandC,VinogradovaY,RobsonJ,MayM,BrindleP:
DerivationandvalidationofQRISK,anewcardiovasculardiseaserisk
scorefortheUnitedKingdom:prospectiveopencohortstudy.BMJ2007,
335:136.
14. RoystonP,CarlinJB,WhiteIR:Multipleimputationofmissingvalues:New
featuresformim.StataJournal2009,9:252-264.
15. PopulationEstimatesbyEthnicGroup:MethodologyPaper.[http://www.
ons.gov.uk/ons/rel/peeg/population-estimates-by-ethnic-group–
experimental-/current-estimates/population-estimates-by-ethnic-group-
Submit your next manuscript to BioMed Central
methodology-paper.pdf],accessed13Jan2012.
16. DowningA,FormanD,ThomasJD,WestRM,LawrenceG,GilthorpeMS: and take full advantage of:
Investigatingtheassociationbetweenethnicityandsurvivalfrombreast
cancerusingroutinelycollectedhealthdata:challengesandpotential • Convenient online submission
solutions[abstract].JournalofEpidemiologyandCommunityHealth2009,
63:88. • Thorough peer review
17. vanBuurenS,BoshuizenHC,KnookDL:Multipleimputationofmissing • No space constraints or color figure charges
bloodpressurecovariatesinsurvivalanalysis.StatisticsinMedicine1999,
• Immediate publication on acceptance
18:681-694.
18. SterneJA,WhiteIR,CarlinJB,SprattM,RoystonP,KenwardMG,WoodAM, • Inclusion in PubMed, CAS, Scopus and Google Scholar
CarpenterJR:Multipleimputationformissingdatainepidemiological • Research which is freely available for redistribution
andclinicalresearch:potentialandpitfalls.BMJ2009,338:b2393.
Submit your manuscript at
www.biomedcentral.com/submit