1504.0 - Methodological News, Sep 2019  
ARCHIVED ISSUE Released at 11:30 AM (CANBERRA TIME) 26/09/2019   
   Page tools: Print Print Page Print all pages in this productPrint All

Introducing an efficient new method for predicting small area values: the Stratified Reweighting Estimator

Knowing the population characteristics in a local area is important for many policy makers. We would like to provide small area estimates as a standard output from our household surveys; however there are two main problems. First, direct survey methods, which use only the local sample in an area, yield very unstable predictions. Second, current methods use tailored statistical modelling methods which, while producing high quality estimates, are also time and resource intensive.

The stratified reweighting estimator (SRE) is designed to enable efficient production of small area predictions for a large range of survey variables.

How the SRE estimator works:

The following steps are applied to a survey unit record file to produce a single large file in which the survey records can appear many times representing different small areas (known as SA2s).

    1. Copy records: Each SA2 belongs to a "stratum" of similar SA2s from across Australia. The survey records from the whole stratum are copied to represent that single SA2 on the output file.
    2. Assign initial weights: The records for an SA2 are assigned weights that add to the SA2 Census count. The weights are larger for units from nearby SA2s (those in the same SA4).
    3. Adjust weights to represent each small area's Census demographics: Records within each SA2 are weighted to add to a range of Census demographic counts. This captures peculiarities of the SA2 as measured at Census time.
    4. Adjust whole file to represent survey estimates: The whole file is finally weighted to reproduce a suite of key survey estimates and demographic totals at Australia level. This is a critical step both in representing the survey time point and in adjusting for the impact of differential non-response.

The resulting weighted file can then be used to generate small area estimates for any survey variable or combination of variables.

Quality assessment for the reweighting approach is based on stability and model goodness-of-fit, and measures of these are provided for each survey variable. Stability is measured by taking the median value (over the small areas) of the Jackknife estimates of the relative standard error (RSE). The model goodness-of-fit is measured by a standardised Wald statistic that estimates how well the SRE predictions fit the survey estimates for each target variable. These two measures, along with knowledge of the predictors used in the model, will enable users to determine if the SRE provides fit-for-purpose predictions for a target variable.

Methodology have used the 2015 Survey of Disability, Aging and Carers to develop and test the SRE, and we are now partnering with our ABS subject matter colleagues to begin implementing the SRE for small area prediction in our household surveys.

For more information, please contact Sean Buttsworth Methodology@abs.gov.au

The ABS Privacy Policy outlines how the ABS will handle any personal information that you provide to us.