CHANGES TO THE RETAIL TRADE SERIES, AUGUST 2007
This issue of Retail Trade, Australia (cat. no. 8501.0, August 2007) introduces some changes to the survey design and methodology for the Retail survey. This is the first survey redesign since the current methodology was introduced in July 2004.
These changes have resulted in revisions to the historical level estimates. The extent of the shift in level varies across the series and, to facilitate comparisons over time, this shift has been smoothed into the full length of the historical series. The impacts on month to month movements have been minimised as far as possible.
These revisions to the historical series have not affected the August seasonally adjusted movements.
WHAT IS INVOLVED IN THE CHANGES
Business Activity Statement information (Total sales) from the Taxation system is used both as a sizing variable for stratification purposes and to form auxiliary information (estimation benchmarks) to support the regression estimation methodology used in the Retail Trade Survey. The utilisation of the taxation data enables the most efficient design for the survey, keeping sample sizes to a minimum while providing accurate results. The Retail Trade Survey design imposes tight sample error constraints at the state/territory and broad industry levels. For other lower level estimates, sample error can be larger. (See paragraphs 32-36 of the Explanatory Notes for more information on the reliability of estimates).
This current redesign utilises updated stratification size values to reallocate the sample, which has resulted in a further reduction in sample size while maintaining current accuracy levels. The redesign also incorporates some improvements in the derivation of estimation benchmarks and improvements in the method of treating survey units that are not representative of other units in the same cohort or stratum (outlier units).
To measure the impact of these changes, a parallel run of the survey was conducted for the month of July 2007. This means information was collected for the old design as well as for the new design allowing old design results to be derived alongside new design results.
WHY THE CHANGES WERE MADE
To avoid survey units moving from one stratum to another, stratification information is generally not updated for each survey cycle, thus reducing the accuracy of the estimates, particularly movement statistics. However, it is now three years since the stratification turnover values have been updated and the increasingly large divergence between the stratification values and the estimation benchmarks (which are updated quarterly) has seen a deterioration in the efficiency of the survey. Survey designers need to strike a balance between keeping strata (or survey cohorts) stable and the overall efficiency of the survey.
HOW THE CHANGES HAVE IMPACTED THE RETAIL SERIES
With new stratification boundaries introduced to align with the refreshed stratification values, a new sample was selected. The new design retains as many as possible of the existing units in sample redistributed on the basis of their refreshed sizing information. However, in addition, a significant number of new units (replacing others rotating out) have been included. Overall, there has been a small ongoing sample reduction of about 200 units.
This rotation and sample redistribution process (coupled with the other minor methodological changes) has had an impact on the level estimates coming from the survey. As this impact is methodological rather than 'real world', to facilitate comparisons over time, the impacts have been measured and smoothed back into the history of the series. This method adjusts historical survey levels while maintaining, as far as possible, the integrity of the month to month seasonally adjusted movements. In a multi-level collection such as Retail Trade, the current backcasting methodology does not enable changes to seasonally adjusted movements to be minimised at all levels of output. In this backcast, level shifts have been estimated and changes to seasonally adjusted movements minimised at the State by industry level. To maintain the additivity of time series at all levels in original terms, higher level aggregate time series are derived from the backcasted State by industry level series. Therefore, the revisions to the historical seasonally adjusted movements of the higher aggregate time series may not necessarily be minimised.
Results from the new survey design for July 2007 indicate an increase of 0.6% in the seasonally adjusted level for the Australian total. Differences at the lower levels vary. The results in this issue reflect the backcast series.
The Retail Trade series will be monitored over the coming months and the backcasting of the historical series may be revised as more data become available on the new basis.
IMPACT ON THE CHAIN VOLUME MEASURES
As historic Retail Trade levels have changed due to the smoothing process described above, there will be an impact on the quarterly chain volume measures. The chain volume measures have been recalculated on the new basis up to June quarter 2007 and are included in this issue (see Tables 14 and 15).