|Page tools: Print Page Print All|
DEMOGRAPHY WORKING PAPER 2004/1 - REVIEW OF INTERSTATE MIGRATION METHOD
In accordance with legislative requirements, the ABS provides quarterly estimates of the population of Australia, the states and territories. These population estimates (known as the estimated resident population, or ERP) are published each quarter for each of the states and territories in Australian Demographic Statistics (ABS cat. no. 3101.0).
The five-yearly Census of Population and Housing provides the benchmark for post-censal population estimates. In order to prepare state level ERP, the ABS first prepares population counts as at the date of the most recent census (i.e. at the time of writing, 7 August 2001) by place of usual residence, adjusting for net under-enumeration in the census and residents temporarily overseas at the time of the census. This census date ERP is backdated to 30 June in the census year by removing the effects of estimated births, deaths and both overseas and interstate migration occurring between 30 June and the census date. Quarterly post-censal ERP is later compiled by updating ERP figures for the previous quarter with estimated natural increase (the excess of births over deaths), net overseas migration and interstate migration (a component of the larger internal migration) for the reference quarter.
Post-censal estimates of births and deaths are calculated using birth and death registrations, while overseas migration is calculated using data collected from passenger cards, visas and passports for those persons travelling into and out of Australia. Data on interstate migration, however, can not be directly estimated. Instead, post-censal estimates of interstate migration are modelled using administrative by-product data (specifically, Medicare changes of address). These modelled estimates are then reviewed against and potentially replaced by the interstate migration estimates from the next census. As such, interstate migration remains the greatest individual source of potential error in state and territory population estimates.
This document describes a recent review of the method used to produce quarterly post-censal estimates of interstate migration. Changes to the method resulting from the review will be applied to interstate migration estimates for September 2001 through June 2006.
STRENGTH OF MEDICARE AS AN INDICATOR
The ABS has evaluated a range of potential sources of administrative data, with Health Insurance Commission data on Medicare changes of address so far found to be the most effective source for estimating interstate migration on a quarterly basis. For more information, refer to ABS Demography Working Paper 2001/5 - Evaluation of Administrative Data Sources for Use in Quarterly Estimation of Internal Migration Between 2001 and 2006.
Medicare card holders are required to register changes of address when they make claims, or when their cards are replaced. However, it is known that some people, particularly younger card holders, do not register changes of address with Medicare, or do so long after the fact. Comparison of Medicare changes of address with census data on the address of respondents one year prior to the census and five years prior to the census suggests that the level of under-reporting in Medicare is fairly constant over time. In addition, this under-reporting seems to be very similar for interstate arrivals and interstate departures, as well as for age and sex groups with low levels of under-reporting.
INTERSTATE MIGRATION METHOD USED FOR 1996 TO 2001
The method used in estimating interstate migration from 1996 to 2001 prior to rebasing, and since 2001 prior to the recent review, is detailed in Demography Working Paper 1999/2 - Estimating Interstate Migration, 1996-2001. This method is summarised below, as the method proposed in the review was essentially an adjustment of the previous method.
For 1996 to 2001, quarterly estimates of net interstate migration were created for the states and territories using Medicare data on changes of address, in conjunction with migration information from the 1996 Census of Population and Housing.
The Medicare data was adjusted by means of expansion factors. The data was broken down by state/territory, sex, single year of age and movement category (arrivals/departures). It was estimated that only those interstate movements of males aged 16-29 years and females aged 18-24 years were under-represented in Medicare data, so expansion factors were only applied to groups with these age-sex characteristics.
Calculating expansion factors
Expansion factors applied to Medicare data were based on the estimated proportion of the population covered. To calculate this proportion, Medicare data for the period October 1995 to September 1996 was compared with 1996 census data on persons' usual residence one year ago. The period of Medicare data chosen reflected the assumption that there is a lag in the registration of changes of address through Medicare.
Expansion factors (calculated as census divided by Medicare data) were estimated for each state or territory, by sex, single year of age and movement category. Analysis showed that expansion factors were greater than 1 (i.e. suggesting 'undercoverage' in Medicare data) for males aged 16-29 years and females aged 18-24 years. It was assumed that Medicare coverage within these age groups could not be less than 50% for any of these groups (i.e. expansion factors were capped at 2). For all other age groups, it was assumed that Medicare data provided a full coverage (i.e. an expansion factor of 1 was used).
Some people will move more than once during a given time period, including some who move to one location, then return to their original location. However, the census only asks for each person's address one year ago, and five years ago, without reference to multiple moves which could have occurred over these one year and five year periods. Census data relating to addresses five years ago are not used in the development of interstate migration estimates, primarily because multiple and return movers become a much greater issue over time.
Since Medicare data do include information on multiple/return movements, these were used to adjust data from the 1996 census on respondents addresses one year ago as follows:
Adjusted Census = Census x (1+proportion of movers in Medicare which were multiple movers).
The proportion of multiple movers in Medicare was estimated by matching individual changes of registration by postcode, sex and date of birth using the September 1995 to September 1996 quarterly Medicare change of registration files.
The use of census data on address one year ago does introduce some potential problems because a relatively short (and therefore potentially more volatile and less representative) time period is used to estimate expansion factors. The smoothing used in generating the expansion factors helps to address these problems.
Two smoothing procedures were used in the production of expansion factors. Firstly, for each state/territory, sex and movement category, data used to calculate expansion factors were smoothed across single years of age, using a three-term moving average. Then, the resulting expansion factors were smoothed, again using a three-term moving average.
Analysis has suggested that registration of changes of address through Medicare generally occurs some time after the actual move. The interstate migration model for 1996-2001 assumed an average registration lag of three months, so Medicare information for a particular quarter was used to estimate interstate migration for the previous quarter. It was not possible to lag this data further (i.e. to assume a delay in registration of more than three months) as this would have impacted on the production and publication of population estimates.
Defence force adjustment
Medicare theoretically covers all Australian usual residents as well as those non-Australian residents granted temporary registration. However, there are a range of Australian usual residents who do not access the Medicare system, primarily due to access to alternative health services. One such group is the military. As such, estimates of interstate migration produced from the interstate migration model described above are adjusted to compensate for defence force movements not covered by Medicare. These adjustments are estimated using counts of defence force personnel by age, sex and state/territory, obtained from the Department of Defence, with 70% of any change in quarterly defence numbers assumed to be due to interstate migration not otherwise covered by the model.
RE-DERIVATION OF INTERSTATE MIGRATION FOR 1996-2001
The model described above for generating post-censal estimates of interstate migration is largely superseded when new census information becomes available. Part of the process of rebasing census ERPs for states and territories is the re-derivation of interstate migration for the intercensal period. The overall approach is to minimise state intercensal discrepancy using census information on one-year and five-year interstate movements. Where this census information does not reduce the intercensal discrepancy, the rebased interstate migration estimates remain largely unchanged from the Medicare-based model. For example, when re-deriving interstate migration estimates for 1996-2001 after the 2001 Census of Population and Housing, estimates for the Northern Territory were not re-derived as this would have increased intercensal discrepancy for the Territory.
Due to the non-compulsory and non-universal nature of the available (indirect) data sources, post-censal quarterly estimates of interstate migration have long been considered the weakest component of population change. For this reason, when new census data are available, interstate migration estimates for the intercensal period are replaced with estimates derived from census data on residence five years ago (for those states or territories where this would reduce intercensal discrepancy). These estimates are then scaled so that they sum to zero at the Australian level. A similar process is carried out for the year prior to the census, using census data on residence one year ago. Again, this is only carried out where it reduces intercensal discrepancy for a state/territory.
The difference between the original interstate migration estimates and the rebased estimates is apportioned across all quarters, movement categories, ages and sexes in the intercensal period in order to minimise quarterly change.
TABLE 1. REBASED AND UNREBASED INTERSTATE MIGRATION ESTIMATES
REVIEW OF INTERSTATE MIGRATION MODEL
The ABS reviewed the interstate migration model in 2003, with the aim of applying a revised model to quarterly estimates for September 2001 to March 2006. The review focussed on the expansion factors applied to Medicare data. The defence force adjustment applied to interstate migration estimates was not reviewed.
A number of scenarios were developed using Medicare data from 1 October 1996 to 30 September 2001, and data from the 2001 census. The scenarios used in the interstate migration review are summarised in the following table and described in more detail under Analysis of results.
TABLE 2. INTERSTATE MIGRATION REVIEW, Scenarios analysed
These scenarios vary in their use of expansion factors, lagging and smoothing, and were each used to estimate June 2001 ERP by state/territory. The extent to which each scenario matched final (i.e. rebased) June 2001 ERP was used as an indication of its accuracy in modelling interstate migration (refer to Assessing Model Quality below). The reasonableness of assumptions in each scenario was also taken into account. Any remaining intercensal discrepancy may be overstated because the new model is closer to the period to which the new model applies (September 2001-March 2006).
Table 3 below shows the intercensal error (both in terms of persons and as a percentage of final 30 June 2001 ERP) for 1996–2001 as well as the estimated intercensal discrepancy produced under each of the review scenarios. The following table summarises the results of this analysis.
TABLE 3. INTERCENSAL ERROR AND ESTIMATED INTERCENSAL DISCREPANCY(a) — 1996 to 2001
Assessing model quality
The alternative interstate migration models (scenarios) were assessed using two main criteria.
It was also desirable (but not essential) that all states and territories shared the same model. This preference stemmed from a concern that models developed for individual states might be effective in reproducing interstate migration for the 1996 to 2001 intercensal period, without adequately describing the underlying relationships between Medicare data and total interstate movers. As a result, such models would not translate well in estimating interstate migration after 30 June 2001.
Expansion factors and lagging
The first scenario considered (S1) assumes that Medicare changes of address perfectly capture all interstate movements (i.e. no undercoverage), and that there is no delay between persons moving address and registering a change with Medicare (i.e. no lagging). These assumptions are known to be false, but are used to provide a baseline for later scenarios. While the results from S1 appeared reasonable for some states when compared with the actual intercensal error, this may be due to errors in interstate arrivals and departures (such as those due to undercoverage and lagging) cancelling each other out.
The next three scenarios (S2-S4) were used to test the impact of lagging and expansion factors. S2 and S4 assumed that there was an average delay of three months (i.e. a lag of one quarter) between a person moving address and this move being registered with the Health Insurance Commission for Medicare purposes. However, S2 assumed that Medicare captured all such movements, while S3 and S4 included expansion factors to adjust for assumed undercoverage of males aged 16-29 years and females aged 18-24 years. The factors applied in these latter two scenarios were referred to as 'raw' expansion factors because they were based on unsmoothed input data from Census and Medicare.
Comparison of the results from S1-S4 indicated that the use of expansion factors and lagging reduced intercensal discrepancy in most states and territories. Later scenarios tested alternative methods of applying expansion factors (smoothing input data and/or smoothing the factors produced), as well as the impact of imposing an upper limit (or 'cap') on the factors.
It was assumed that much of the variability in data from one single-year age group to the next was due to 'random noise', so smoothing was incorporated in some form in each of the remaining scenarios analysed. A simple three-term moving average was applied in several different ways: to smooth outputs (i.e. expansion factors); to smooth inputs (i.e. Census and Medicare data); and to smooth both the inputs and outputs.
Overall, smoothing of expansion factors produced only small improvements in the estimated levels of intercensal discrepancy; in some cases, results were worsened when the factors were smoothed. Greater improvements (represented by smaller intercensal discrepancies) were gained by smoothing Medicare and Census data used to produce the expansion factors. For most states and territories, the best results were found in those scenarios from which expansion factors were created using smoothed input data.
Smoothing applied to ACT data
While smoothed input data improved results for all the states and the Northern Territory, it increased the estimated level of intercensal discrepancy in the ACT. Data from the 2001 Census on respondents' places of usual residence one year ago shows a sharp difference between the numbers of interstate arrivals for males aged 17 years and for males aged 18 years (and likewise for females). It appears that these movements (likely to be associated with students arriving in the ACT for tertiary study) are not being picked up through Medicare. The use of a 3-term moving average to smooth these data reduces the peak and results in fewer estimated arrivals to the ACT after expansion factors are applied, thus producing higher estimated intercensal discrepancy.
A different smoothing method was applied in S10 to address this issue. Expansion factors in this scenario were based on Census data that had been smoothed with a 3-term moving average with, for interstate arrivals of males and females (but not interstate departures), a number of Census data points set such that the peak at age 18 was retained. Expansion factors for the remaining states and territories were set as for S6.
In the interstate migration model used for 1996-2001, expansion factors calculated as being greater than 2 (i.e. less than 50% coverage estimated for Medicare data) were set to 2. The rationale for 'capping' expansion factors was that this would reduce the influence of outlying extreme results, such as unusually low registrations in particular age/sex groups.
Capping was applied to expansion factors in S7 and S9, but did not produce improvements in estimated intercensal discrepancy. Also, capped expansion factors produced slightly increased levels of estimated intercensal discrepancy for the Northern Territory and the ACT.
Results from scenarios 1-10 indicated that the preferred interstate model should contain the following components:
This preferred model was reflected in scenario 10 (S10) and provided the best overall results. The ABS intends to apply this model to produce interstate migration estimates for September 2001 to June 2006. A defence force adjustment will continue to be applied to estimates produced by this model to compensate for movements of defence force personnel not covered by Medicare.
For further information on the interstate migration method, contact Rhonda de Vos on Canberra (02) 6252 6639, email <firstname.lastname@example.org>.
ATTACHMENT 1: ERP AND COMPONENTS OF CHANGE, 1996 to 2001
Table A1 below contains preliminary (unrebased) and final (rebased) ERPs and components of population change for the 1996 to 2001 intercensal period.
TABLE A1. ERP AND COMPONENTS OF CHANGE, 1996 to 2001
These documents will be presented in a new window.