The ABS has been working to meet strong demand for new and improved regional labour statistics. Our primary focus has been enhancing monthly measures of employment, unemployment, and related indicators at the Statistical Area Level 4 (SA4), using small area modelling techniques. To achieve this, the ABS implemented a Rao-Yu model which uses administrative data sources such as Single Touch Payroll (STP), JobSeeker, and Youth Allowance to help smooth design-based SA4 estimates from the Labour Force Survey (LFS) (see [1,2]). These modelled estimates were integrated into the monthly Labour Force, Detailed publication from the March 2024 reference period [3]. Secondary goals include producing SA4 statistics disaggregated by age and sex, and assessing the feasibility of generating estimates at finer geographic levels, such as SA3 and SA2.
Design-based estimates for SA4-age-sex groups have been published in Labour Force, Detailed since 2006. However, these estimates are highly volatile due to small sample sizes within each SA4-age-sex category. To improve the stability and quality of these estimates, we modelled the unweighted sample counts \(x_{dt}\) for each labour force status, age group, and sex category using a binomial distribution:
\[x_{dt}∼Binom(n_{dt}, p_{dt} )\]where denotes an SA4 and a month; is the sample count, and is the proportion in the labour force category. The resulting hierarchical logistic regression model is :
\(log(p_{dt}/(1-p_{dt} ))=βX_{dt}+v_d+u_{gt}\)
Here, \(v_d\) is a normally distributed zero mean random intercept at SA4 level, \(u_{gt}\) an AR(1) process defined over broader geographic regions, and \(X_{dt}\) the design matrix for the main effects components which consist of metro/ex-metro specific intercepts and linear trends. New South Wales, Victoria, and Queensland are each split into metropolitan and ex-metropolitan regions; the Australian Capital Territory is combined with metropolitan New South Wales; and the SA4s within the remaining states and NT are similarly grouped into metropolitan and ex-metropolitan categories.
We estimate \(p_{dt}\) using the R package mcmcsae [4], then multiply by the estimated resident population to produce initial counts. These are then post-stratified to maintain consistency with the corresponding SA4 estimates. The modelled SA4-age-sex statistics are significantly less volatile than the direct estimates, and are now included in the Labour Force, Detailed from the July 2025 reference period.
We are currently investigating the feasibility of producing labour force statistics at geographies below SA4. High-quality predictors are essential at these finer levels. STP data is a key input for deriving employment estimates, and ABS is investigating methods to efficiently pre-process this data, in particular the use of horseshoe priors to adjust for trend breaks in the STP time series while maintaining sparsity in regression coefficients. These trend breaks are artefacts of administrative processing within financial years rather than reflecting underlying labour market conditions.
Monthly JobSeeker recipient data is also available at fine geographic levels. ABS is examining methods to account for the dynamic relationship between JobSeeker and unemployment, enabling JobSeeker data to support unemployment estimates at SA3 and SA2 levels.
If these approaches prove effective in efficiently pre-processing administrative information at SA3 and SA2 levels, they may also provide a pathway to improving the stability of SA4-age-sex administrative data.This could allow domain-level administrative data to be incorporated into future iterations of the SA4-age-sex model, further enhancing the quality of regional labour market outputs.
For further information please contact Mark Ioppolo.
References
[3] Labour Force, Australia, Detailed | Australian Bureau of Statistics
[4] mcmcsae: Markov Chain Monte Carlo Small Area Estimation