Fellows Research Meetings

Statistical impact measurement with state space modelling approach

Oksana Honchar
Australian Bureau of Statistics

One of the most common risks related to the implementation of a major statistical business transformation is that – despite improvements in data collection efficiency, process methodology and overall data quality the induced statistical impacts of measurement could be mis-interpreted as real world change. In this research a methodology is proposed for applying a seemly unrelated time series equation (SUTSE) model in state space form to historical data of multiple relevant data sources that measure a similar concept to the target survey variable but are not subject to the measurement change. The statistical impact can then be assessed by intervention analysis taking advantage of the cross correlations and leading properties between the target survey variable and the multiple data sources. The method takes into account real world changes as reflected in the multiple data sources, and therefore augments an intervention analysis relying on the survey data alone. In this research, historical data from the Australian Labour Force Survey (LFS) is used to test the methodology.

Using historical data, we also present a state space modelling approach for assessing statistical impact of the web-form introduction in Australian LFS. A multivariate state space model was implemented at the rotation group level data to take into account the web-form introduction scheme and to take advantage of sampling error correlation due to the rotating panel design. This state space modelling approach can be easily extended for a parallel run design where the ‘old’ and ‘new’ survey designs are run in parallel for a period of time in order to collect information to enable the quantification any statistical impact.

Last reviewed: 7 September, 2017