The ABS has used estimated response propensity, calculated at address level, to improve the representativeness of the final responding sample to the 2024 Time Use Survey.
Traditionally, the ABS has used expected response rates at geographical levels to ensure sufficient selections are made to deliver the required responding sample. For example, if 1000 responses are required from New South Wales and 1000 from Victoria, and the anticipated response rates are 80% and 75% respectively, then 1250 selections are made in New South Wales and 1333 in Victoria. This will deliver the required sample size in each state, so long as the anticipated response rates are achieved. Despite this, the sample may still be unrepresentative within each state. For example, if single person households are less likely to respond than family households, then the sample in each state will over-represent family households and under-represent single person households. This imbalance can be corrected during weighting, although increased variation in the size of weights leads to reduced accuracy in the final survey estimates.
An alternative approach adopted for the 2024 survey is to predict the likelihood that each address on the sampling frame will respond to the survey. The sampling frame is linked to the Person Level Integrated Data Asset (PLIDA), this process attaches administrative data to each address on the frame. These administrative data are then used to generate the estimates of response propensity for each address. Once the propensities have been produced, they are then released in decile categories, and the average propensity for each category is used as the anticipated response rate for the addresses in each category.
The usual process is then used to inflate the required number of responses by dividing by the anticipated response rate to give the number of selections required in each decile category, and consequently a selection probability for each address. This means the previous method of making a geographically-based allowance for non-response has been replaced by a method of making an address-level adjustment. If the estimated response propensities accurately predict the final achieved response propensities, then the final responding sample will be representative not just across geographical areas, but also by household characteristics (as reflected in the administrative data).
An analysis of response rates at the end of data collection shows reduced variation in achieved responding sample sizes by response propensity decile, and so reduced variation in the estimation weights. The reduction in variance has been assessed as equivalent to a 15% increase in sample size.