• Summary
• Downloads
• Explanatory Notes
• Related Information
• Past Releases

PREDICTING SURVEY ESTIMATES BY STATE SPACE MODELS USING MULTIPLE DATA SOURCES

The ABS is embarking on a transformation program, which includes, amongst other things, re-engineering and consolidating collections, using different collection modes for survey data, and using different, but more efficient, sampling frames and estimation methods for official statistics. Whilst this transformation is expected to bring about positive changes to official statistics, there is also a risk that such changes could lead to impacts on some ABS time series. The challenge for the ABS is to develop methodologies to monitor, measure and, where needed, adjust for any such impacts.

The methodology the ABS is proposing to implement for this purpose makes use of the data of related series. A special multiple time series model called a Seemingly Unrelated Time Series Equation (SUTSE) model has been investigated as a basis for predicting a target survey estimate using multiple data sources.

Where related series measure a similar concept to the target survey variable, but are not subject to measurement change, these can assist in understanding the change that occurs on the target survey variable. Under this method, the statistical impact can be assessed by intervention analysis, taking advantage of the cross-correlations and leading properties between the target survey variable and the other related series. The power of this method has been tested by estimating historical supplementary survey effects and the effect of past questionnaire redesign using Australian Labour Force Survey data. This work has also been extended in a number of other directions.

A case study involving LFS unemployment confirmed that a standard bivariate SUTSE model with claimant count data offered improvements in terms of prediction error, detecting outliers and structural changes in the target unemployment estimates.

However, available related data sources may not have appropriate properties for applying a standard SUTSE model to predict survey estimates efficiently. As part of these investigations the ABS developed a strategy to select valuable data sources and adjust the way a SUTSE model is applied to take advantage of SUTSE modelling strength. Another case study considered employment estimates from the LFS, and demonstrated that such a strategy also has the potential to work much better than a univariate structural time series model, by borrowing strength from multiple source data in an efficient way.

Further Information
For more information, please contact Mark Zhang or Oksana Honchar (methodology@abs.gov.au).

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