Page 314 - Special Topic Session (STS) - Volume 3
P. 314
STS547 Owen A.
within-household non-response. The variance of the estimator was also
higher, reflecting its simpler form.
Abbott et al (2015) explored the weighting class approach further in the
context of a census, showing some of the empirical properties in this context
when compared to a standard DSE approach.
ONS (2017) reports on additional work exploring weighting classes when
some activity data provides an improved administrative data base. However,
it was not possible to separate out the biases due to failures in some of the
other assumptions underpinning the method. This made it very difficult to
assess whether the approach was ‘better’ than a standard DSE.
In summary, this work highlighted that the weighting-class approach is
analogous to a DSE where the assumption of zero within-household non-
response means that the estimator is no longer as susceptible to individual
level over-coverage on the first source. It is still susceptible to over-coverage
on the frame and over-coverage on the survey, but both of these are deemed
less of a risk. Whilst it is an attractive approach due to its over-coverage
properties, like DSE, it would require a second estimation process to adjust for
biases due to failure of assumptions. In this case it would be for within-
household non-response.
2.3 Other methods
There are other flavours of dual-system estimation being explored for the
purposes of estimating population size using administrative data. Again, these
approaches essentially trade off the underlying assumptions to attempt to
reduce bias, choices being made based on the context. Here we describe the
work being undertaken by two National Statistical Institutes in this area,
although these are not exclusive.
2.4 Trimmed DSE
The Central Statistics Office Ireland is exploring a trimmed DSE approach
(Dunne, 2018). The idea behind this approach is heavily based on having good
quality activity type information (either directly from a source or via linkage
across multiple sources) to deal with over-coverage. The activity information
is used to trim out records from the administrative data, calculating a new DSE
along the way. If the records being trimmed are genuine over-count, the DSE
will decrease (as the records will be in the ‘admin only’ cell of the DSE and thus
be inflating the estimate). Trimming continues until the DSE stabilises and/or
the variance grows too much, and at that point the DSE is taken as the best
estimate.
This approach makes a lot of sense – it is turning an administrative source
from one with over-coverage (and undercoverage) into a source with only
under-coverage. The DSE methodology works much better under those
conditions, so it is a clever way of getting one assumption to fit much better
and thus reduce bias. However, it does perhaps make the estimator more
303 | I S I W S C 2 0 1 9