|Page tools: Print Page Print All|
FEATURE ARTICLE 2: RECASTING 20 YEARS OF ERP
WHY RECASTING WAS NECESSARY
After each Census, the ABS rebases ERP to the latest Census count. As part of this process, the ABS adjusts ERP to take into account any intercensal error by distributing the error evenly across each quarter over the five year period since the previous Census.
Preliminary rebasing of ERP after the 2011 Census followed this standard methodology, with the estimates published on 20 June 2012. However, the standard rebasing treatment could not credibly account for the large intercensal error between 2006 and 2011, and the resulting ERP series showed implausible growth for this period. The large intercensal error identified was predominantly due to a change in methodology in the 2011 Census Post Enumeration Survey (PES).
The PES is conducted after each Census in order to assess coverage of Census counts, as represented by the key measure of net undercount. The 2011 PES utilised the new methodology of Automated Data Linking (ADL), which resulted in better linking and matching of PES and Census records, and a better measure of net undercount.
A statistical impact study was conducted to determine the impact of using the new ADL method, and it was found that the 2011 net undercount was approximately 40% lower than it would have been if the previous methods had been used. The ABS estimates that the previous net undercount would have been substantially lower for all previous post enumeration surveys, had this methodology been available, and the large intercensal error in 2011 is therefore likely to have been accumulated over a period greater than the usual five years. It was decided that this large statistical impact (which accounts for around 84% of the 2006-2011 intercensal error) should be incorporated into ERP by distributing it over a period longer than the usual five years.
For more information, please refer to:
THE DECISION MAKING PROCESS
The ABS’ intention to recast ERP data over a period longer than the usual 5 year rebasing period was first proposed in the 27 September 2012 release of Australian Demographic Statistics, along with an open invitation for comment. After an extensive consultation process, the ABS made the decision to recast ERP over a period of 20 years. This was announced in the 18 December 2012 release of the same publication.
The 20 year period (reflecting four intercensal periods) was decided on as the period of time that would result in an estimate of population growth that reflected the growth observed in the historical data for population components (that is, births, deaths and migration), which are the best data source for measuring population change over time.
Recasting over the standard 5 years or over a 10 year period were both options that were considered, but deemed sub optimal, as the resulting ERP series would have low growth rates that would be more reflective of the change in methodology, rather than reflect the best estimate of growth over the period. Recasting over a longer period (such as back to the beginning of ERP in 1971) was shown to achieve very little statistical gain.
METHODOLOGY USED IN RECASTING
The processes and methods used to recast the data were developed and quality assured by a team of demographic and methodological specialists within the ABS. These methods were guided by a series of principles that were developed during the consultation process, and published in the 18 December 2012 release of Australian Demographic Statistics. These principles established that:
2. The use of ADL in the PES has been a major improvement in how the ABS measures Census coverage, and the 2011 net undercount should be used to inform historical understanding of Census coverage.
3. Population growth for the 2006-2011 period should, as closely as possible, reflect the growth in the population components (i.e. births, deaths and migration) for all spatial levels (i.e. national, state, and sub-state).
4. Any assumptions should be based upon the best available data.
5. Any revision to the historical ERP series should maintain the demographically plausible relationships between the fundamental building blocks of population series (e.g. age-sex profiles).
6. Where revised data exist for population components data, they should be used regardless of whether they were available at the time of previous rebasing processes (e.g. recently identified late registration births for NSW for 2005-2010 should be included).
Census Point Adjustments
The usual rebasing treatment of intercensal error distributes the error evenly over the five year period, as there is no further data to inform upon the distribution of the error within the period. Rather than distributing the impact of ADL back evenly over the 20 year period, the recasting process differentially adjusts each of the four intercensal periods, based on all information available for that period. This resulted in a greater impact on the data around the 2006 Census point, gradually decreasing to a minimal impact on the 1991-1996 data.
The recasting process involved calculating revised undercount adjustments at each Census rebasing point (ERP at 30 June of each Census year- 1996, 2001 and 2006), with these points then used as the base population for the rebasing of quarterly intercensal estimates between these recast base points. This recasting process therefore only involves change in the ERP series. No adjustment was made to the actual Census counts.
Adjustment calculations followed a top-down approach; the first step was to calculate the total adjustment to undercount at a national level for each Census point, referred to as the 'magnitude'. This was then apportioned to the states and territories, followed by lower levels of geography, with age profiles calculated and applied following the same top-down sequence.
The magnitude of the 2006 adjustment was taken directly from the ADL Statistical Impact Study (SIS) results (i.e. 246,900).
For 2001, the reduction was based on the 2006 reduction and an offsetting impact of a change in PES methodology that was made between the 2001 and 2006 post enumeration surveys. The 2006 PES implemented an improved estimator, known as Prediction Regression (PREG). At the time of the 2006 PES, the impact of the change was 108,400, but this was not considered significant enough to warrant recasting of the historical ERP data. However, the information on the impact of PREG was incorporated into the recasting revisions, in accordance with Guiding Principle 4. For more information on the 2006 change in PES estimation, see Research Paper: An Estimating Equation Approach to Census Coverage Adjustment, May 2007 (cat. no. 1351.0.55.019) and Census of Population and Housing Details of Undercount, Aug 2006 (cat. no. 2940.0).
For the 1996 adjustment, an adjustment was derived based on the need to minimise the change between the previously published estimates and recast estimates over the period 1991-2001, and which would maintain growth rates over the 1991-1996 and 1996-2001 periods as closely as possible.
Apportioning to the State and Territory Level
The apportionment of the national adjustment to the state and territory level was calculated based on a combination of intercensal error and the 2006 PES adjustment. The use of these two sources reflects the strength of each source - the PES estimate directly links the adjustment to the original state or territory allocation of net undercount, while component growth and related error is considered to be a reliable indicator of changing state and territory distribution over time. This use of intercensal error also enabled the process to directly address an observable degree of implausible variability in state and territory level intercensal error over time.
To derive the state and territory level adjustments for the 2001 Census point, the 2006 state and territory split was offset by the state and territory specific impact of PREG. The differential impact of PREG on the various states and territories, together with addressing the relatively consistent pattern of positive intercensal error for Western Australia, Tasmania and the two territories in recent rebasing cycles, explain why there are positive value adjustments for some jurisdictions. The use of a composite factor method ensured that the most plausible adjustments were derived and applied to population estimates for each state and territory, according to the best available information.
The 2001 state and territory split was then multiplied by the ratio between the 1996 and 2001 magnitude adjustments to derive the 1996 state and territory split.
While this publication only includes recast ERP at the Australia and state and territory levels, ERP is also available at lower levels of geography, and sub- state recasting is also being undertaken. These data will be available on 30 August 2013, when it is published in the next release of Regional Population Growth, Australia (cat. no. 3218.0).
Adjustments for lower levels of geography involve apportioning the state level adjustments to the capital city and rest of state level, followed by further apportionment of the adjustment among the component sub-state regions. These adjustments also take into account the age and sex structure of each region.
Age and Sex Distribution
The age structure for each adjustment was adopted directly from the ADL statistical impact study results, grouped into the age groups of 0-14, 15-29, 30-49 and over 50. The use of the larger age groups partly mitigated the high standard errors associated with the estimates of impact, but also reflected the relative consistency of the impact across many of the 5 year age groups. The greatest adjustment was required for young adults (15 to 29 years), reflecting both the relatively high undercount in each PES, and the higher impact of ADL on this age group. This age group is traditionally the most difficult to enumerate in the Census, in part due to its mobility.
The existing sex ratio was maintained at the state and territory level for each five year age group.
RECASTING QUARTERLY ESTIMATES
With the adjusted population levels set for each Census point, the quarterly intercensal estimates were then recast progressively from the 1991-1996 period to the 2006-2011 period, according to the standard rebasing method. This method assumes that the intercensal error has accumulated by an equal number each quarter over the five-year intercensal period. For example, the intercensal error at the adjusted 1996 Census point is allocated evenly across the 1991-1996 period, with the population at the end of the first quarter adjusted by 1/20 of the (revised) intercensal error, the second quarter adjusted by 2/20, the third by 3/20, and so on. For further information on the rebasing process, please refer to Population Estimates: Concepts, Sources and Methods, 2009 (cat. no. 3228.0.55.001).
IMPACT ON THE DATA AND ITS USE
When comparing the original ERP series with the recast series, the proportionality of age groups, states and territories and sexes are not significantly changed.
The following population pyramids show the relative similarity in the population distribution by age before and after recasting, noting the largest changes were for the 15-29 year old age groups, given the higher undercount for this age group. These population pyramids are for 2006, which was the Census point most significantly affected by recasting.
In addition, while previous ERP levels have been lowered, growth rates remain relatively consistent with those published before 2011 rebasing.
Due to the impact of ADL on the estimation of the 2011 intercensal error, the preliminary rebasing of ERP resulted in an implausible growth rate for the 2006-2011 intercensal period, compared to the rest of the series. For more information, see Technical Note: The Impact of Improvements to the 2011 PES on Measuring Population Growth, Australian Demographic Statistics, Dec 2011. By recasting the data, the 2011 intercensal error (84% of which reflected the PES methodological improvement) has been distributed across a longer time period. This means that recasting has rectified this implausibility, restoring the growth rate to a similar level to that observed prior to rebasing.
It is expected that in the future, ERP will continue to be rebased on a five yearly basis after each Census. In contrast, recasting should be seen as an exceptional event made necessary by the significant methodological improvement in 2011; it is not anticipated that the recasting process will be repeated in relation to ERP in the foreseeable future. In addition, the experience gained from the recasting process and its impact on ERP will be used to inform plans for the 2016 rebasing and beyond.
Data that uses ERP as an input will generally be impacted by recasting in the same way that data would be impacted by the usual 5-yearly rebasing process. Demographic data will be revised to reflect the updated ERP figures, according to the timetable below.
ABS survey data will not be updated, with the exception of the Labour Force Survey, which is rebenchmarked to be consistent with rebased ERP, as standard practice after each rebasing cycle. Revised Labour Force Survey estimates for 1991 onwards will become available with the release of January 2014 data in February 2014. National Accounts data will also be revised to incorporate the recast ERP. Revisions will be published with the release of June 2013 National Accounts data in September 2013.
FOR MORE INFORMATION
For more information, please contact Phil Browning on (02) 6252 6639, or email email@example.com.
These documents will be presented in a new window.