|Page tools: Print Page Print All|
Feature Article - Technical report: Labour Force Survey sample redesign
In less populated areas, an additional stage precedes the selection of collection districts to ensure that the sample is not too geographically spread (as that would lead to unacceptable enumeration costs).
The sample of non-private dwellings is obtained by compiling a list of non-private dwellings in Australia. A systematic random sample is taken from this list (with probability proportional to size) in such a way that each region across Australia and each different type of non-private dwelling is represented. For smaller non-private dwellings, each occupant is included in the survey; for larger dwellings, a sub-sample of occupants is taken.
Allocation of sample
The LFS is designed primarily to provide reliable estimates of key labour force aggregates for the whole of Australia and, secondarily, for each state and territory.
The most accurate national estimates would be obtained if the total sample for Australia were allocated in proportion to the population of each state or territory. However, for each state or territory to have estimates as accurate as one another, equal size samples would be needed for each.
The allocation of the sample across the states and territories is a compromise between one that would be optimum for national purposes (i.e. the same sampling rate in each state and territory) and one that would give each state and territory the same accuracy (i.e. the same sample size in each). That is, the proportion of the population in the sample (known as the sampling fraction) differs across states and territories, but not to the extent that would realise identical sample sizes for each state and territory. Within each state and territory, each dwelling has the same probability of selection.
One of the primary requirements of the survey is to provide a measure of change in the characteristics of the labour force over time, especially month-to-month variations.
The best estimate of month-to-month change would require data to be collected from essentially the same sample of dwellings each month (while providing for population growth). However, it is not reasonable to retain respondents in the survey indefinitely. A proportion of the sample is therefore deliberately replaced each month. This procedure is known as sample rotation.
Since the monthly LFS began in 1978, one-eighth of the sample has generally been replaced each month. The sample can be thought of as consisting of eight sub-samples (or rotation groups), with a new rotation group being introduced into the sample each month to replace an outgoing rotation group. This replacement sample usually comes from the same area as the outgoing one.
Sample rotation enables reliable measures of monthly change in labour force statistics to be compiled, as seven-eighths of the sample from one month is retained for the next month's survey. At the same time, the sample rotation procedure ensures that no dwelling is retained in the sample for more than eight months.
The component of the sample that is common from one month to the next makes it possible to match the characteristics of most of the people in those dwellings: this group is referred to as the 'matched sample'. The availability of this matched sample permits the production of estimates of 'gross flows' - the number of people who change labour force status between successive months.
2001 SAMPLE REDESIGN
Reflecting its importance in maintaining the efficiency and effectiveness of the LFS, development of the 2001 redesign included the following key aims:
A number of improvements were considered in developing the new design. The more significant changes being implemented are:
Use of a constant sampling fraction between sample redesigns has the effect that the number of dwellings in the sample increases as the population grows.
The graph below shows the number of persons enumerated in the LFS sample each month from 1992 to 2002, illustrating the gradual increase in the number of people enumerated between each redesign. While this results in some improvement in the accuracy of the survey results, the improvement is partially offset by a deterioration in the efficiency of the sample in the period since the previous redesign.
Further, as more dwellings are added to the survey over time, the operational costs of collecting the data increase. To offset these increases in cost, the sample size is reduced at each redesign. The decrease in sample size following the 1991 and 1996 Census redesigns can be seen on the graph below. The grey line at the right shows the expected decrease in sample size during the period November 2002 to June 2003, as the sample from the 2001 Census redesign is implemented.
Following implementation of the 2001 design, the initial sample size is expected to be about 3% smaller than at the start of the 1996 design. Despite this reduction in sample size, the levels of sampling variability (averaged over the life of the new sample) associated with estimates of both level and month-to-month movement are expected to be little different from those realised over the life of the previous design. This is the result of a small gain in efficiency in the 2001 design compared with the previous design.
When fully implemented in June 2003, it is expected that there will be about 28,600 private dwellings and 1,900 non-private dwellings in the sample each month, representing about 1 in 224 (0.45%) of dwellings across Australia. This is expected to result in about 60,000 people responding to the survey each month.
Unlike previous designs, the state and territory sampling fractions were an output from the design process, rather than an input.
Sampling fractions have changed little beyond that which would be expected from adjusting the 1996 design for population changes, except for the Northern Territory. Greater efficiency gains were found in the Northern Territory sample under this redesign, enabling a smaller sample to be allocated. As a result, a substantial improvement in accuracy of unemployment estimates has been realised for the Northern Territory, partially offset by slight reduction in accuracy of employment estimates.
The following table gives the sampling fractions used for each state and territory, from the 1976 Census redesign to the new, 2001 Census redesign.
REDESIGN SAMPLING FRACTIONS
Relative standard errors
Averaged over the life of the new sample design, relative standard errors (RSEs) for employment and unemployment at the national level are expected to be the same as those achieved under the previous sample design, as the table below shows.
RSEs for employment and unemployment at the state or territory level are expected to be practically the same as those achieved under the previous sample design. In the case of the Northern Territory, a slight increase is expected for the RSE of employment, more than offset by an improvement in the RSE for unemployment.
While the redesign results in a smaller sample, the improved design more than offsets the increase in variance that would result from a sample size decrease taken in isolation from the redesign.
LFS RELATIVE STANDARD ERRORS
Further information about sampling variability and standard errors for LFS data will be published later in 2003 in Information Paper: Labour Force Survey Standard Errors (cat. no. 6298.0).
In order to reduce the potential impact of the change in sample on labour force statistics, the new sample is being introduced progressively, taking advantage of the existing sample rotation scheme.
The private dwelling sample in larger urban centres and less remote areas, representing just over four-fifths (82.1%) of the total sample, is being phased-in over the period November 2002 to June 2003. Within these areas, one-eighth of the new sample will be introduced each month under existing sample rotation arrangements.
The rest of the sample (in the more remote, less populated areas and for non-private dwellings) was introduced in two stages: in November 2002 for New South Wales, Victoria, Tasmania, the Northern Territory and the Australian Capital Territory; and in December 2002 for Queensland, South Australia and Western Australia.
This method of implementation means that most of the changes to labour force statistics due to differences between the two samples, or any other influences, will be spread over the eight months. This approach is broadly comparable with that adopted for the 1996 redesign. In contrast, the approach adopted for the 1981 redesign saw the new sample introduced in one month, while in the 1986 and 1991 redesigns, the new sample was introduced over four months.
To assist in the interpretation of published estimates during the phase-in period, the ABS has been investigating the labour force characteristics of the new sample in comparison with the sample that it is replacing. Analyses have included using alternative estimation methods that place greater emphasis on the sample that is common between consecutive months of the survey. These methods have produced estimates of employment growth for the five-month period November 2002 to March 2003 that are very similar to the published increase, although with a less volatile monthly pattern. For more detail, see Notes on Estimates published in recent issues of Labour Force, Australia, Preliminary (cat. no. 6202.0).
Standard errors associated with the redesigned sample (when fully implemented) will be similar to those of the previous sample, as discussed above. However, standard errors of monthly movement during implementation of the new sample will be higher than for other periods, because month-to-month correlations of survey estimates are reduced. This arises because new-sample dwellings rotated into the sample each month tend to come from different areas to the old-sample dwellings rotated out (by contrast, in periods other than when a new sample is being implemented, dwellings rotating into the sample each month tend to be next door to those rotating out).
During the implementation period, movement standard errors will be highest for October-November 2002 and November-December 2002, because of the higher sample rotation rates in these two periods for the more remote areas and non-private dwellings.
While additional sampling error is expected during implementation of the new sample, there is no evidence from analyses undertaken so far that the change of sample has otherwise affected estimates of employment growth. Analysis is continuing, and the ABS will publish results when they are completed.
For further information about the Labour Force Survey sample design, or about the statistical regions defined from November 2002, see Information Paper: Labour Force Survey Sample Design (cat. no. 6269.0). Descriptions of the underlying concepts and structure of Australia’s labour force statistics, and of the sources and methods used in compiling the estimates are presented in Labour Statistics: Concepts, Sources and Methods (cat. no. 6102.0), which is also available on the ABS web site (About Statistics - Concepts and Classifications).
These documents will be presented in a new window.