Page 323 - Special Topic Session (STS) - Volume 3
P. 323
STS547 John D. et al.
target population records missed by both the list and the sample into the
target population and to impute an overcoverage indicator for the list records
in the observed (0, 1) cell. Records imputed as overcoverage are excluded from
population estimates. The second approach weights the list records by the
ratio = ̂ 11 + ̂ 10 which is the ratio of one minus the over-coverage
̂ 11 + ̂ 01
probability to one-minus the undercoverage probability. In practice, the
weights are specific to particular covariate combinations. Using either the
imputation or weighting approach sub-group population estimates can be
readily obtained, either by counting records or summing weights within the
sub-groups of interest. To represent uncertainty these calculations are
repeated for each draw from sample from the posterior of the cell
probabilities.
The key assumptions underpinning this approach are:
• the sample is drawn from the target population with known inclusion
probabilities for each person (in reality the sample inclusion
probabilities may be estimated)
• selection in the sample is conditionally independent of inclusion on the
list, given the covariates included in the model
• linkage between the sample and the SPD is done in an error free way
• there is no misclassification with respect to covariate information
The key innovation in this approach is that it does not require sampling
directly from the SPD to estimate overcoverage, an idea first proposed by
Zhang (2015).
Graham and Lin (2019) acknowledge there is still work to be undertaken
with this approach particularly in areas related to clustering in sample design,
record linkage error, misclassification in data and outline a number of
directions that this work could take.
3. Discussion and Conclusion
The situation in Ireland and New Zealand are similar. Both countries
compile an SPD that may contain overcoverage and undercoverage. In the
case of Ireland, the approach is to compile an SPD that only has
undercoverage and then focus on statistical methods to adjust for
undercoverage which may be sizable. In New Zealand, the approach is
different in that the SPD attempts to be as close as possible to the target
population and the statistical methods are then required to only make minor
adjustments for undercoverage and overcoverage. However, both methods
can be deployed in scenarios where both undercoverage and overcoverage
exist.
Both methods require a second data source to adjust for coverage error in
the SPD. Henceforth, we refer to this second list as the coverage list. In both
cases, the coverage list must come from the target population with no
312 | I S I W S C 2 0 1 9