GENERALISED REGRESSION VARIANCE ESTIMATION FOR ABS ECONOMIC SURVEYS
The Australian Bureau of Statistics (ABS) is investigating improving the efficiency of its estimation methodology for its economic surveys by using Business Activity Statements (BAS) that are submitted to the Australian Taxation Office by Australian businesses. Improving efficiency will allow the ABS to reduce the sample size of its surveys, and consequently, the cost of producing publication estimates, while maintaining their level of accuracy. Statistical agencies around the world are facing similar challenges, as administrative data are becoming accessible.
General Regression (GREG) is a statistical estimation methodology that may use several items of auxiliary information, such as the various items available in BAS data. GREG estimation is potentially significantly more efficient than the alternatives, which are to use either no auxiliary information or a single auxiliary item as in ratio estimation. The ABS has an estimation system, called GREGWT, that implements the GREG estimation methodology.
While the ABS can calculate GREG estimates for its economic surveys by using GREGWT, it must also be able to estimate the precision, or variance, of these estimates. Special Projects Unit in the Statistical Services Branch is drafting a working paper evaluating methodologies for estimating GREG variances, including: Bootstrap, Jacknife and Balanced Repeated Replication. This evaluation includes theoretical criteria as well as the ability to meet ABS' output requirements.
The working paper recommends the Bootstrap for calculating GREG variance estimates for ABS economic surveys. It develops the theory supporting the use of Bootstrap to calculate variances of level and rate estimates and their movements between two time points, in single-phase and two-phase sampling schemes.
For more information, please contact James Chipperfield on (02) 6252 7301.