South Australian Branch SSA Meeting: March 2019
The speaker for the March meeting of the SA Branch was Claire Clarke, a methodolgist at the Australian Bureau of Statistics (ABS). Claire gave an engaging talk on the methods being developed at the ABS to improve coordination of samples selected for its household and business surveys.
The ABS conducts a wide variety of sample surveys on social and economic topics. Since different ABS surveys often sample from the same population, sample selection needs to be coordinated in order to manage the survey load placed on individual households and businesses. Also, sample coordination does not always seek to minimise overlap: for surveys which produce estimates every month or every quarter, it is desirable to have high overlap between the samples selected in successive periods. Under the design-based sampling framework used for ABS surveys, a requirement of the sample coordination method is that it preserves each unit’s selection probability specified in the sample design.
For many years the ABS has applied the ‘synchronised selection’ method to achieve sample coordination of its business surveys. Claire explained the method is not particularly flexible, and so the goals of sample coordination are not achieved when units change strata or there are extensive changes to the structure of the sampling frame.
The general ‘Conditional Selection’ method which the ABS has been developing largely addresses the limitations of synchronised selection. Under Conditional Selection, units belonging to the same sampling stratum are selected with different probabilities according to their history of selection across previous surveys. For example, assuming it is desirable to minimise the extent of overlap between the sample for an upcoming survey and the samples of previous surveys, for the upcoming survey the units in the population which have not been previously selected will have highest conditional probability. Claire used an example to explain the calculation of the conditional probability. She illustrated how the probability associated with each potential selection history can be used to ensure each unit is selected with the desired unconditional probability.
In the second half of her talk Claire discussed practical issues for implementation. One such issue is controlling the sample size within each stratum. If it is necessary to control the sample size in each stratum, the selection method needs to be adapted and it is not possible to precisely preserve the desired unconditional probabilities. Another issue is managing the selection history data. Although the selection history (and associated probability the history) must be tracked for every unit in the population, the storage requirements are manageable because the number of possible histories for each unit increases linearly with the number of prior surveys being tracked.
The ABS is in the process of adopting the Conditional Selection method, and it has already been applied for selection for some ABS household surveys.
By Julian Whiting