imputation in the 2001 census
DESCRIPTION
Imputation in the 2001 Census. Robert Beatty NILS User Forum 11 December 2009. Coverage. How Census deals with Missing households Missing people within households Incomplete returns. Coverage. Census is statutory Census Act (Northern Ireland) 1969 Penalties for non-compliance - PowerPoint PPT PresentationTRANSCRIPT
Imputation in the 2001 CensusImputation in the 2001 Census
Robert Beatty
NILS User Forum
11 December 2009
CoverageCoverage
How Census deals withMissing householdsMissing people within householdsIncomplete returns
CoverageCoverage
Census is statutoryCensus Act (Northern Ireland) 1969Penalties for non-complianceTherefore counts everyoneDoesn’t it?
CoverageCoverage
Population in thousands
Published Census figure
MYE
1991 1,578 1,607
CoverageCoverage
Population in thousands
Published Census figure
MYE
1991 1,578
(enumerated)
1,607
(best estimate)
Coverage - internationalCoverage - international
Australia 2006 – 96% coverageDon’t impute but adjust MYEsNew Zealand 2006 – 95% response rateNZ imputed for non-response, but only on 4
key variablesCanada ‘adjust for non-responding
households’ – need to know about occupied households
Adjustment issuesAdjustment issues
1991 coverage – 98%But inference about population?Non-response not homogeneousYoung adultsLower social classDeprived areas
Coverage - 2001Coverage - 2001
Acknowledge under-enumeration1991 Census 1,578k MYE 1,607kDecision to adjust Census 2001 databaseObjective – all Census outputs to fully
reflect whole population‘One Number Census’Census = MYE
CoverageCoverage
Population in thousands
Published Census figure
MYE
1991 1,578
(enumerated)
1,607
2001 1,685
(adjusted)
1,689
Coverage - 2001Coverage - 2001
‘One Number Census’ method Basic principle to use a large-scale Census
Coverage Survey (CCS) to estimate under-enumeration in sampled areas
Apply survey estimates elsewhere
Census Coverage SurveyCensus Coverage Survey
UK split into about 100 Estimation Areas (each about 0.5m population)
Three in Northern IrelandAbout 200 postcodes / 3,000 households per
Estimation AreaThree socio-economic strata within EASeparate analysis in each strata within EA
Census Coverage SurveyCensus Coverage Survey
Fieldwork about 3 weeks after Census dayFace to face interviewsTrained interviewersGiven map of postcode boundaryAsked to re-enumerate the postcodeShort questionnaire - coverage
MatchingMatching
Forms scanned into systemSpecial matching software developedDatabase retrieval systemCCS returns carefully matched with Census
returns – error rate estimated to be under 0.1 per cent
Dual System Estimator (DSE)Dual System Estimator (DSE) Use matched Census and CCS data DSE estimates adjustment for those missed in both
Census and CCSCounted By CCS
Yes No
Counted Yes n11 n10 n1+
By Census No n01 n00 n0+
n+1 n+0 n++
DSE estimate for the area (under certain assumptions):
n++ = n1+ n+1 n11
DSE : Simple ExampleDSE : Simple ExampleFish pondFish pond
Day 1: Catch 950 fish, mark with a red dot. Day 2: Catch 900 fish, mark with a blue dot. Matched: 855 had blue and red dots. Question – how many fish in the pond?
Dual System Estimator (DSE)Dual System Estimator (DSE)
Counted Day 2
Yes No
Counted Yes 855 95 950
Day 1 No 45 n00 n0+
900 n+0 n++
DSE estimate of the actual number of fish:
n++ = 950 900 855 = 1,000
AnalysisAnalysis
Separately for each age-sex group, within each stratum, within each EA
Apply DSE method to each sampling point (postcodes) within CCS area
Estimate function DSE = f(observed count)Apply to all other sampling points within
stratum (within EA), and aggregate
Ratio EstimationRatio Estimation
Regression-type estimator
Each dot represents a CCS area
Use Census figure to estimate “true” figure
Census
DSE
The One Number Census The One Number Census processprocess
CENSUS
CENSUS + CCS
ESTIMATE BY AGE AND SEX
FOR EA
EA ESTIMATES
ADJUSTED INDIVIDUAL AND
HOUSEHOLD DATA AND TABLES
NI POPULATION ESTIMATE
MATCHING
Quality Assurance
Dual System and regression estimation
CCS
QA
Imputation controlled to EA
estimates
Sum
Imputing householdsImputing households
Use dummy forms as locationUse dummy forms as ‘constraint’?Dependence on enumeratorsIreland 2006 – 15% of properties vacant
One Number Census outcomeOne Number Census outcome
2001 Census response rate of 95%4.3% in wholly imputed households (mostly
linked to dummy forms(3.0%))0.4% additional people in already
enumerated householdsImputed 80,000 people
CoverageCoverage
Population in thousands
Published Census figure
MYE
1991 1,578
(enumerated)
1,607
2001 1,685
(adjusted)
1,689
Distribution of LAD Level Underenumeration
0
10
20
30
40
50
60
70
80
Response Rate
Fre
qu
en
cy
Response rates by ageResponse rates by age
80
85
90
95
100
'0-4
'10-
14
'20-
24
'30-
34
'40-
44
'50-
54
'60-
64
'70-
74
'80-
84
age
male
female
Quality of returnsQuality of returns
So far, considered non-respondentsPerson & Household imputationWhat about quality of returns actually
made?Decision taken to go for ‘complete’ returnsItem imputation
Edit and Impute - EditEdit and Impute - Edit
Limited number of ‘hard’ edits – can’t be married if aged under 16
Larger number of ‘soft’ edits - quality
Edit and Impute - ImputeEdit and Impute - Impute
General principle of ‘complete’ data setNo ‘Not stated’ entries in outputsItem imputation usedDonor imputation systemNo different in principle to systems used in
sample surveys
Edit and Impute - ImputeEdit and Impute - Impute
Level of item imputation differed by variable
Not applied to religion
SummarySummary
Objective in 2001 that Census outputs should reflect whole population
Person and household imputation5% of persons imputedComplete records generated for all returns
through ‘item’ imputation
I told them in 1951 it was just you, me and the dog, but they keep coming back every 10 years to check.
Looking forwardLooking forward
Date for your diaries …27 March 2011
Any questions?Any questions?
Usual residence definitionUsual residence definition
Historical – present on nightMost countries now ‘usually resident’Definitions do exist (UN)2001 – self-assessed2011 – instructions‘Intention to stay’