1
Measuring Uncertainty inPopulation Estimatesat Local Authority Level
Ruth Fulton, Bex Newell, Dorothee Schneider
2
Outline
• Project aim
• Overall method
• Method internal migration
• Method international migration
• Outputs
3
Project aim
Improve understanding, measurement and reporting of the quality of population estimates at LA level
• Obtain overall quality measures for annual
population estimates at LA level
4
Mid-Year Population Estimates
• Cohort component method
Pop.(t) = pop.(t-1) + births
+ internal net migration
• Determining associated uncertainty is complexMixed sources: Census, administrative sources, surveys
Different estimation methods
+ international net migration
– deaths
5
Measuring uncertainty: Overall method
- Components with biggest impact:2001 Census-based estimate
Internal migration
International migration
- Estimate distribution of error for component - Combine error estimates into overall quality measure for MYE at LA level
Error (t) = error (t-1)
+ error (net internal migration)
+ error (net international migration)
6
Internal migration
• Estimates based on GP registration data • Sources of uncertainty in estimates related to:
• Migrants missing from GP register
• Time lags between moving and re-registration
• Double counting of school boarders
7
Method for internal migration
• Benchmark approach
• Uses adjusted 2001 Census data as benchmark
• Applies model from 2001 to subsequent years
• Limitation – does not cover all quality issues
8
Method for internal migration (ctd.)
• Movers in Census: those with other address one year ago
• Movers in PRDS: those with different addresses in two downloadsCensus data adjusted to be as similar to PRDS data
as possible
• Compare observed number of migrants to a ‘true’ number of migrants
• Error represented by scaling factor of truth (Census)/PRDS
9
Age pattern
• log(Census/PRDS)• shows double counting of school boarders• shows undercount of young male migrants
Mean log scaling factors for inflows by age and sex
-0.4
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0 10 20 30 40 50 60 70 80
Mea
n ls
f
Males Females
Age
10
Geographical variation
- Scaling factors vary by area
- Undercount in urban areas or areas with high proportion of students
- Cluster analysis
Mean log(Scaling Factors)Inflows
11
Model
-.5
0.5
1M
ean
pre
dict
ed
log(
sf)
0 20 40 60age
group 1 group 2
group 3 group 4
Mean predicted log scaling factors, Inflows male
• Fit model to log of scaling factors of groups of LAs• Obtain predicted values and residuals
Error measure is obtained by simulating from this distribution
12
Distribution of estimated inflows
05.
0e-0
4.0
01.0
015
.002
Den
sity
6000 6500 7000 7500 8000estimated 'true' flow
green: PRDS flow, red: Census (observed truth)example LA
13
International migration
• Focuses on intentions-based IPS estimates• Multi-stage approach to distribute to national estimates
to lower levels of geography
IPS direct estimate
Calibrate to Labour Force Survey (LFS) 3 year average
Distribute using IPS 3 year average
Distribute using immigration model
National
Regional
Intermediate (NMGi)
Local Authority (LA)
14
• Produce error distribution for statistical error
Bootstrapping approach
Resampling IPS
Resampling LFS (regional level)
Reproduce estimation method with new samples
International migration (ctd)
15
Outputs
Outputs• Composite quality measure will be derived from the
overall error distribution• LAs will be banded based on this measure