|Page tools: Print Page Print All RSS Search this Product|
Feature Article - Technical report: New Labour Force Survey sample selections - analysis of the effect on estimates
CHANGE IN SEASONALLY ADJUSTED EMPLOYMENT ESTIMATES, November 2002 to June 2003
The phase-in of the new LFS sample was undertaken with an overriding objective of maintaining continuity of time series for LFS estimates. The ABS undertook a number of investigations to evaluate whether the phasing-in of the new sample had an impact on LFS estimates. These investigations included:
The ABS also conducted analyses using two alternative estimation methods that place greater emphasis on the sample that is common between consecutive months of the survey.
In conclusion, these analyses and investigations indicated that the implementation of the new sample had no major effects on the continuity of LFS estimates. As expected, the relative standard errors of estimates of monthly movement increased during the phase-in period. The analysis of the characteristics of incoming and outgoing rotation groups suggests how this increased sample variability might have manifested itself in practice each month. Both alternative estimation methods produced estimates of change in employment over the eight month phase-in period that are similar to the published figure, although with a less volatile pattern. While increased volatility is unavoidable during a phase-in period, these studies have suggested some changes to operational arrangements which may enable the ABS to control elements of that volatility a little better in the post-2006 Census redesign.
ANALYSIS OF OPERATIONAL PROCEDURES
ABS investigations of the possible impact of operational procedures on LFS estimates took a broad focus and included issues relating to: sample design; sample size; survey scope or definitions; response rates; selection; derivation and weighting programs; interviewers; coverage of non-private dwellings; and field and office procedures.
These investigations found no errors that could have contributed to the high employment growth during the phase-in period. There were a number of improvements made to the operational procedures used for the new sample, but these did not have a significant impact on the survey estimates.
SAMPLING VARIABILITY DURING PHASE-IN PERIOD
The ABS analysed sampling variability during the phase-in period of the new sample, focussing on standard errors for estimates of monthly change of employed, unemployed and not in the labour force.
The average sampling variability of level estimates (as measured by standard errors) associated with the new LFS sample design were expected to be broadly comparable to that of the previous sample. The same is true for movement estimates, once the phase-in is complete. However, standard errors of monthly movements were expected to be higher during the sample phase-in period, particularly for the months of November and December 2002.
In the LFS, households are interviewed each month for eight months, with 1/8th of the sample being replaced each month. The replacement (or rotation) of a sample is usually achieved by selecting another cluster of dwellings, generally in close proximity to the dwellings leaving the sample. As a result, the incoming rotation group is likely to possess similar socioeconomic and labour force characteristics to the outgoing rotation group.
The main difference during a phase-in period is that the outgoing rotation groups are not replaced by nearby dwellings. The incoming rotation groups are therefore more likely to exhibit different labour force characteristics to outgoing rotation groups while the new sample is being introduced.
There is no systematic difference between the outgoing and incoming groups. They are both 'random probability' samples of the Australian population, and they both create unbiased estimates of labour force characteristics.
The ABS estimated that the standard errors on estimates of monthly movements during the phase-in of the new sample, compared to estimates based on the old sample, would be increased by 5% for employed persons, 1% for unemployed persons and 3% for persons not in the labour force. The increase in standard errors was expected to be higher still for movements in November and December, as a result of the higher proportion of the new sample being introduced in November and December.
ANALYSIS OF CENSUS RESULTS
Sampling variability occurs because one group of respondents is randomly selected instead of another. It is possible to gain some insight into the specific differences between the outgoing 1996 sample and the incoming 2001 sample by using Collection District (CD) data from the 2001 Census of Population and Housing. While this does not provide information on the differences that existed during the phase-in period, it does shed some light on the types of differences that existed between those CDs in 2001.
For example, the following table shows the employment to population ratio (EPR) for August 2001, for the CDs in the incoming rotation groups and for the CDs in the outgoing rotation groups, for each of the eight months of the sample phase-in period. Differences between the estimated EPRs for the incoming and outgoing rotation groups will contribute to sampling variability.
ESTIMATED EMPLOYMENT TO POPULATION RATIO (a), Collection Districts (b), of outgoing and incoming LFS rotation groups, August 2001
The EPR was chosen as a measure to enable comparison of the level of employment between groups of different sizes. The EPRs for the CDs in incoming rotation groups were higher than those in outgoing rotation groups for the first three months (i.e. in November, December and January), lower for the following two months, higher again for the sixth month and quite similar for the last two months. Since the rotating sample moved, in the first three months, to geographic areas with higher EPRs at the time of the 2001 Census, there is a possibility that these rotation groups contributed to the growth in the LFS employment estimates in these months.
ALTERNATIVE ESTIMATION METHODS
Another way of examining the impact of the sample change is to use the information reported by respondents during the phase-in period, but use alternative estimators that are less sensitive to the sample rotation effects. Note that, while these estimators produce more stable estimates of monthly movement in situations like this, they have other properties that make them less attractive in the usual situations.
The two alternative estimation methods used (matched rotation group, and composite estimation) place greater emphasis on the sample that is common between consecutive months of the survey. See Note on Methodology at the end of this paper for more detail about these estimators.
One drawback to the matched rotation group method is that over time the resulting level estimates drift away from the true level. In addition, the seasonal adjustment process used for both of these analyses was only approximate because the seasonal adjustment was applied to 'hybrid' series. These 'hybrid' series were formed by appending the matched rotation group series (or composite estimation series) for November 2002 to June 2003, to the published employment series for the period to October 2002.
The following graph compares the seasonally adjusted matched rotation group and composite estimation employment series with the published series. The cumulative change in employment over the eight months of the sample phase-in for both alternative methods is similar to the published figure, although the matched rotation group series (and, to a lesser extent, the composite estimation series) has a less volatile monthly pattern than the published series.
IMPLICATIONS FOR 2006 LFS SAMPLE PHASE-IN
Alternative matching of rotation groups
Rotating in groups from different geographic areas during the phase-in period can have an impact on the sampling variability of the estimates. This is particularly the case when rotation groups with relatively high EPR are replaced by groups with relatively low EPRs, and vice versa.
This phenomenon can be alleviated somewhat by matching outgoing rotation groups with incoming rotation groups with similar EPRs (as measured in the most recent census), and thereby reducing the volatility in survey estimates. The graph below shows how a matching of incoming and outgoing rotation groups could have been used to reduce the differences between the EPRs of these groups.
Investigations for the phase-in of the 2006 LFS sample redesign will look at more closely matching outgoing rotation groups with incoming rotation groups, based on employment characteristics from the 2006 census. This may lead to smaller differences between the old and new samples, and to a smaller contribution of the sample phase-in to LFS sampling variability.
Timing of phase-in period
The 2001 sample redesign was phased in over a period of eight months from November 2002 to June 2003, with around 1/8th of the new private dwelling sample being introduced each month under existing sample rotation arrangements. This reduces the likelihood of dramatic movements directly due to the new sample phase-in, as well as providing operational advantages and helping to minimise sampling error. This strategy, and the timing, will be reviewed for the 2006 sample redesign, drawing upon experiences gained during the 2001 phase-in.
For further information about the analyses in this article please contact Bill Allen on Canberra 02 6252 6302, or via email at email@example.com. For information about the Labour Force Survey please contact Craig Blair on Canberra 02 6252 6565, or via email at firstname.lastname@example.org.
NOTE ON METHODOLOGY
Matched rotation group method
The matched rotation group method calculates monthly movements using the 7/8ths of selected dwellings that are common between consecutive months under the LFS rotation scheme (i.e. the incoming group of dwellings is dropped from the current month's estimate, and the outgoing group is dropped from the previous month's estimate).
The matched rotation group series shown in this paper has been derived by adding successive estimates of monthly movement (beginning with November 2002) to the published level estimate for October 2002, and then seasonally adjusting the whole series (as if it were a continuous time series). Hence the seasonally adjusted series is only approximate, because the matched rotation group estimates may have different seasonal characteristics from the published estimates. Also, analysis has shown that there is a small systematic difference in responses between the outgoing and incoming groups of dwellings in the LFS, related to the length of time that respondents have been in the survey. This results in a small bias in the matched rotation group estimates of monthly movement, which will accumulate over time as successive monthly movements are added to the employment estimate for October 2002.
In addition to accumulating bias, the matched rotation group method has higher standard errors than the published series, for estimates of employment change over periods longer than six months.
These deficiencies render the matched rotation group method inappropriate for other than short-term comparisons.
Composite estimation method
The composite estimation method calculates an unbiased estimate of the employment level for any given month. It gives most weight to the dwellings in common with the previous month, but also gives some weight to the outgoing and incoming dwellings. The composite estimation series shown in this paper has been derived by combining original estimates for the period November 2002 to May 2003 with published (original) estimates for the period to October 2002, and then seasonally adjusting the whole series (as if it were a continuous time series). Hence, although the composite estimation method is sound, the seasonally adjusted series is only approximate, because the composite estimates may have slightly different seasonal characteristics from the published estimates.
Further details of the matched rotation group and composite estimation methods can be found in an article in the May 1998 issue of Australian Economic Indicators (cat. no. 1350.0).
These documents will be presented in a new window.