Towards a Fully Adjusted Census Database for the 2011 Census Christine Sexton (ONS) Alan Taylor (ONS) James Brown (ADMIN @ IoE)
Jan 09, 2016
Towards a Fully Adjusted Census Database for the 2011 Census
Christine Sexton (ONS)Alan Taylor (ONS)
James Brown (ADMIN @ IoE)
Outline
• Overview of the Census Coverage Assessment and Adjustment Strategy for 2011
• The 2001 Adjustment Strategy
• Learning from 2001
• Assessment of the 2001 Adjustment System
• The Way Forward
Overview of the Coverage Assessment and Adjustment Process
Estimation
Matching
Adjustment
2011 Census
Quality Assurance
Census Coverage
Survey
The 2001 Adjustment Strategy
• Stage 1: Imputation of missed households (with people)
─ Model to derive predicted census household coverage probabilities using matched census to CCS data to obtain coverage weights─ tenure, ethnicity, household structure
─ Calibrate coverage weights to key variable estimates─ tenure exactly
─ Impute households with people into the database─ whole household records copied
The 2001 Adjustment Strategy
• Stage 2: Imputation of missed individuals into counted households
─ Model to derive predicted person within counted census household coverage probabilities using matched census to CCS data
─ age, sex, activity, household structure, LA
─ Calibrate coverage weights to key variable estimates at local authority level ─ age-sex groups exactly
─ Impute people into census counted households─ whole person records copied
The 2001 Adjustment Strategy
• Stage 3: Final adjustment
─ Further adjustments to meet local authority level estimates for age-sex groups and household size distributions─ taking out imputed individuals─ putting in extra individuals(pruning and grafting)
Ref: Steele, Brown and Chambers (2001), JRSS, series A.
• Insufficient control of household size and characteristics for imputed households─ Too many people in certain age-sex groups added at household imputation stage
─ Much time spent “pruning and grafting”
• Insufficient heterogeneity in the imputed population for some characteristics─ Whole records copied to imputed households and individuals
─ Ensured Census edit rules satisfied but may not reflect variability in population
Learning From 2001
Assessing the Performance of the 2001 System
• Used simulations
– Uses 2001 census extracts as the ‘true population’
– modelled 2001 matched census and CCS data– 10 simulated censuses and CCSs for one Estimation Area (two LAs)
– Census coverage 94%– 200,000 households– 490,000 persons– Used true totals as calibration constraints
LA age-sex group totals, activity, tenure, household size
Performance measures
100
T
TT10N
1
RAB
N
1e
10
1ie
(adj)ei
100
T
TT10N
1
RRAMSE
N
1e
10
1i
2
e(adj)ei
Relative Average Bias Results for Tenure
-16-14-12-10-8-6-4-20
Tenure
Rel
Ave B
ias
(%)
Census
Adjusted
RRAMSE Results for Tenure
010
203040
5060
Tenure
RR
AM
SE
(%
)
Census
Adjusted
Relative Average Bias Results for Males by Age
-15
-10
-5
0
5
10
Age group (males)
Rel
ativ
e A
vera
ge
Bia
s (%
)
Census Adjusted
RRAMSE Results for Males by Age
0
5
10
15
20
25
30
01-
45-
9
10-1
4
15-1
9
20-2
4
25-2
9
30-3
4
35-3
9
40-4
4
45-4
9
50-5
4
55-5
9
60-6
4
65-6
9
70-7
4
75-7
980
+
Age group (males)
RR
AM
SE
(%
)
Census Adjusted
Relative Average Bias Results for Activity
-20
-15
-10
-5
0
5
Activity
Rel
Ave
Bia
s (%
)
Census
Adjusted
RRAMSE Results for Activity
0
20
4060
80
100
120
Activity
RR
AM
SE
(%
)
Census
Adjusted
The Way Forward
• Aim to improve imputation by gaining better control of numbers of individuals imputed into households and their characteristics
─ Correct distribution of age group and household size at lower levels of geography
─ Reduce time spent on final adjustment
(pruning and grafting)
Modelling Missed Individuals
• In 2001 we modelled individuals missed within counted households─ no direct control of individuals missed within missed
households
• Proposed new model –
all missed individuals in single model ─ missed within counted households ─ missed within missed households
• Calibrate coverage weights for all individuals
then split weights into two components based on the model
Reverse the order of imputation
• In 2001 household imputation carried out first─ Within household imputation used to make up
shortfall─ Household weights did not match individual totals─ Imputed households did not contain correct types
of individuals
Reverse the order of imputation
• New person model gives direct control over split between two sources of undercount─ Can put missed individuals into counted households
first to complete counted households
• Then model census household coverage
• Calibrate household weights to key variables at EA level – tenure and household size
• Also calibrate household weights to key individual level variables from the persons in missed HHs totals – age-sex groups – at LA level to recover totals at the individual level
Conclusions
• By implementing the proposed changes we aim to improve on the 2001 system by gaining better control of the age-sex by household size distribution of the adjusted database and reduce the need for the final stage adjustment
• Analysis of 2001 method gives us a bench-mark to compare changes
• Work in progress
Questions?