IMPROVED ESTIMATION METHOD
In June 2007, the ABS will introduce an improved method of estimation for the Labour Force Survey (LFS). The new method, known as composite estimation, is more efficient than the current estimator. That is, the composite estimator achieves a given level of standard error at lower cost than the current estimator.
The new estimation method will be introduced with the release of May 2007 labour force statistics on 7 June 2007 in Labour Force, Australia (cat. no. 6202.0). At the same time, the ABS will release revised historical LFS statistics based on the new estimation method, back to April 2001. An updated standard error model will also be introduced to reflect the composite estimation method.
The change in estimation method will have an impact on all LFS products. Detailed information on statistical impacts will be provided in future releases of this publication and in an information paper to be released in late April 2007.
REVIEW OF CURRENT ESTIMATION METHOD
The current LFS estimator derives estimates of the number of people employed, unemployed and not in the labour force by applying expansion factors (or weights) to the LFS sample responses for the reference month so they add up to independent estimates of the civilian population aged 15 years and over (called population benchmarks). The benchmarks are classified by geographic area, age and sex.
The ABS has been investigating composite estimation methods for several years. Some of the ABS research findings are available in the article 'Can Labour Force Estimates be Improved using Matched Sample Estimates?' in the May 1998 issue of Australian Economic Indicators (cat. no. 1350.0), available free from the ABS web site.
NEW ESTIMATION METHOD
The composite estimation method being implemented is a modified version of a Best Linear Unbiased Estimator. The new composite estimator will combine data collected in the previous six months with the current month's data to produce the current month's estimates. Technical details about the method can be found in Research Paper: The impact of rotation patterns and composite estimation on survey outcomes, MAC Paper, 1998 (cat. no. 1352.0.55.017). The method is similar to that being used for labour force surveys in several other countries.
In the LFS, dwellings remain in the survey for eight consecutive months, with one-eighth of the sample being replaced each month. This means there is a seven-eighth overlap in the dwelling samples in adjacent months, a six-eighth overlap in the samples two months apart, and so on. The composite estimator exploits the high correlation between overlapping samples across the current and immediately preceding months to achieve lower standard errors than the current estimator.
EFFECT ON THE LEVELS OF LABOUR FORCE ESTIMATES
The new composite estimator produces estimates of employment and unemployment which are slightly lower on average than those produced by the current estimator. This effect applies for original, seasonally adjusted and trend series. However, the underlying directions of the series are unchanged under composite estimation.
Analysis of labour force data for the period April 2001 to January 2007 shows that, for seasonally adjusted series at the Australia level, employment estimates are 0.07% lower on average under composite estimation than under the current estimator. Unemployment estimates are 1.60% lower, whilst the unemployment rate is 0.08 percentage points lower (on average). The participation rate is 0.10 percentage points lower on average.
The tendency for slightly lower levels of employment and unemployment under composite estimation is due to the 'time in survey' effect which has long been observed in the LFS. This effect refers to the tendency of the incoming one-eighth sub-sample each month to produce slightly higher employment and unemployment results (and lower numbers of not in the labour force) than the other sub-samples that have been in sample for some months. The composite estimator changes the impact of the 'time in survey' effect on survey estimates because it puts less weight on the dwellings that are new in the sample. See the April 1998 release of the publication Working Papers in Econometrics and Applied Statistics, No. 98/2 (cat. no. 1351.0).