Page 313 - Special Topic Session (STS) - Volume 3
P. 313
STS547 Owen A.
for the 2021 Census. Brown et al (2018) provide a description of DSE and its
underpinning assumptions, which will not be reproduced here.
In the context of using administrative data instead of a Census to derive
population size estimates, there will still be a requirement for a second ‘count’
of the population which is used for estimating coverage (and hence size). In
the UK, there are plans to develop a Population Coverage Survey (PCS) which
will play that role, much like the equivalent Census Coverage Survey (CCS). The
PCS will borrow many sample design elements from the CCS – it will be a short
survey focusing on counting households and people, around 300,000
households per annum.
Using the PCS within a DSE framework seems like the obvious answer, but
it is susceptible to bias due to over-coverage in either source, and as noted
previously many of the administrative lists in the UK suffer from this. In
addition this would also require high quality linkage between the
administrative data and the survey, which is achievable but costly.
2.2 Weighting class
An alternative to DSE is a weighting-class approach (see Lohr, 1999). The
basic idea is to use information known about household responders and non-
responders to estimate response rates within classes. In the context of deriving
population size from administrative data, the known information are the
administrative data entries linked to the address frame being used by ONS for
all of its statistical processes. The PCS sampled addresses are linked to the
frame so that responding and non-responding addresses are identified.
The key attractions of this approach (compared to DSE) are that high
quality individual level linkage is no longer required as the only requirement
is linkage to the address frame, and the estimator is less susceptible to over-
coverage in the source used to estimate the weights (in this case the
administrative data). The reduction in the matching requirement is extremely
attractive, as it is a costly and time intensive undertaking to ensure that
matching is completed to a high standard.
However, the approach does have some drawbacks. It makes no allowance
for persons missed in addresses captured by the survey, which is generally
about 2 to 3 per cent in previous census coverage surveys. In addition, it can
also be biased in the same way as DSE as it also assumes homogeneity of
response probability within classes.
As part of the initial work to explore estimation options for using
administrative data in place of a census for population size estimation, ONS
(2013) reported on the work undertaken to apply such an approach. This
assumed a population coverage survey as described previously was designed
and available. Such a survey was simulated from the 2011 Census data, along
with administrative data. The weighting class approach performed as
expected, being much less sensitive to over-coverage but biased due to survey
302 | I S I W S C 2 0 1 9