MULTI-STAGE CONDITIONAL SELECTION FOR HOUSEHOLD SURVEYS
The ABS has recently moved to the use of an address register for selecting samples for household surveys. This has required some changes to the way household survey samples are selected and managed. Previously, broad geographic areas (First Stage Units or FSUs) were selected every five years, using Karmel or Keyfitz selection to avoid overlap with broad regions from previous selections. Overlap and rotation were managed within these areas by a fairly manual process of ordering Base Frame Units or BFUs (small geographic areas) and clusters of dwellings, and field staff counting through them. With the address register we will be selecting dwellings directly, although to facilitate the hiring of local interviewers and minimise interviewer travel we will still be doing this within the framework of BFUs and FSUs. This means that the manual approach for overlap and rotation control is no longer viable.
Conditional selection is a new method for controlling overlap and rotation for sample surveys that allows the usage of individual selection units to be tracked over time. It avoids selecting units that are undesirable (perhaps because they have been selected before) by assigning them a low 'conditional' probability of selection. This is done in such a way as to guarantee the sample is selected without bias and maintaining the desired probability of selection overall. Conditional selection is a very powerful and flexible method for overlap and rotation control. In particular, it imposes no limits on survey stratum boundaries or on how probabilities of selection are calculated. However, in its basic form conditional selection only applies to a single stage of selection. The ABS has been working on extending conditional selection to apply to multi-stage selection processes, and in particular the multi-stage approach used in ABS household surveys.
In order to extend conditional selection to two or more stages it is necessary to reflect the conditional selection probabilities of earlier stages of selection in the second or subsequent stages. This is relatively straightforward. Logistically, it is generally preferable for sample enumeration to continue sampling in an area for as long as there are dwellings available to select. As such, it is also necessary to introduce usage measures. The usage measures count sample usage within areas and prompt conditional selection to make an area undesirable for further selection once it no longer has enough sample. However, a difficulty arises in ensuring that areas get a fair chance of selection in later samples. Because we stay in an area until it is fully used, there is a possibility that the sample will get 'trapped' in larger BFUs and never move on to smaller BFUs, effectively altering their probability of selection. To avoid this we introduce a new generalisation of conditional selection called Multiple Outcome Conditional Selection. This approach allows the sample the possibility of moving out of a BFU and returning later, thus allowing all BFUs a fair chance of selection at any point.
Another issue with multi-stage selections taken over a long period of time is that the geographical areas involved change as the population grows and evolves. We have developed approaches for splitting and/or combining selection histories and recalculating usage measures when geographic boundaries are redrawn.
The first ABS surveys to have their samples selected using conditional selection were the Labour Force Survey, the Survey of Disability and Carers and the National Aboriginal and Torres Strait Islander Health Survey. These began going into the field in July 2018.
For more information, please contact Claire Clarke Claire.Clarke@ABS.gov.au