3125.0 - Demography Working Paper 2001/3 - Improving Household Estimates, 2001  
ARCHIVED ISSUE Released at 11:30 AM (CANBERRA TIME) 06/06/2001   
   Page tools: Print Print Page Print all pages in this productPrint All

Introduction

1. Estimates of the number of resident households are produced quarterly by the ABS. The estimates are based on the household population series (that is, the estimated resident population (ERP) that usually lives in private dwellings), to which propensities to form households are applied. While national household estimates have been reliable, the data have been volatile at the State, Part of State, and household composition levels. Household composition refers to the combinations of persons aged 0 to 14 with persons aged 15 or more. In response to requests to improve this series, the ABS has improved the method of household estimation. This paper examines the underlying data problems, and demonstrates how the new series is an improvement on the previous method.

Summary of results

2. The previous method employed by the ABS was criticised for two key reasons: firstly, the series underwent a "kink" following each census, as census estimates of the number of households were used for the censal points. The improved series of household estimates smooths in the census figure, so that points before and after the census quarter estimate are a weighted average of the census estimate and the household population-propensity based estimate.

FIGURE 1 : ESTIMATED RESIDENT HOUSEHOLDS - Australia, as at 30 June
FIGURE 1 : ESTIMATED RESIDENT HOUSEHOLDS - Australia, as at 30 June



3. The impact of the new method is clearly noticeable at the national level, and is more pronounced at the sub-national level, as illustrated by the data from the Northern Territory.

FIGURE 2 : ESTIMATED RESIDENT HOUSEHOLDS - Northern Territory, as at 30 June
FIGURE 2 : ESTIMATED RESIDENT HOUSEHOLDS - Northern Territory, as at 30 June



4. The second key criticism of the previous method of producing household estimates was that the annual estimates were volatile, particularly for sub-national estimates and for estimates by household composition. The reason for this variability in the data is that the propensity to live in a particular household composition within a certain area is measured by the Labour Force Survey, and as such is subject to sampling error. This is particularly the case when we are considering small cells, such as household composition in Hobart, for example. The improved method of household estimation tries to reduce this sampling error by two mechanisms; firstly, the household composition propensities are applied at a broader level than they were under the previous method, and secondly, the period-to-period variability is reduced through application of a smoothing filter.

FIGURE 3 : ESTIMATED RESIDENT HOUSEHOLDS - Hobart, as at 30 June
FIGURE 3 : ESTIMATED RESIDENT HOUSEHOLDS - Hobart, as at 30 June


FIGURE 4 : ESTIMATED RESIDENT HOUSEHOLDS - Adelaide, as at 30 June
FIGURE 4 : ESTIMATED RESIDENT HOUSEHOLDS - Adelaide, as at 30 June


FIGURE 5 : ESTIMATED RESIDENT HOUSEHOLDS - Balance of Victoria, as at 30 June
FIGURE 5 : ESTIMATED RESIDENT HOUSEHOLDS - Balance of Victoria, as at 30 June



Changes to the Household Estimation Method

5. There were three changes implemented to improve the method used to estimate the number of usually resident households within Australia. The underlying propensity method, used before the modifications detailed here, is described in Household Estimates, 1986, 1991-94 (Cat. No. 3229.0).

6. Firstly, the level of aggregation used before undertaking propensity calculations was increased. This meant that the cells used for estimation from the Labour Force Survey were broader than under the previous method. Specifically, the previous method applied a household composition propensity to ERP with the finest level of disaggregation being "four or more persons aged 15 or more living with three or more persons aged 0 to 14". For many geographic areas (eg, Northern Territory, Tasmania) the sample from the Labour Force Survey was especially small, meaning that the resulting household estimate was volatile. The new method now being employed has "three or more persons aged 15 or more living with two or more persons aged 0 to 14" as the finest level of disaggregation.

7. This improvement substantially reduced the variability of the data, but the resulting series was still considered too volatile. Specifically, whilst there were far fewer large changes in quarterly estimates, the changes that remained were still not reasonably explained by changes in ERP; there is an implicit assumption here that living arrangement propensities do not change substantially from quarter to quarter for a large geographic region. The changes in the data that were being observed were still felt to be being driven by the sampling error from the Labour Force Survey, since with collapsing of living arrangement/household size groupings the sample sizes from Labour Force Survey were still quite small. In response to this observed variability a filter function was introduced in an attempt to smooth out the quarterly volatility, whilst still allowing ongoing trends to be measured. After several filter functions were trialled, a 13 term Henderson filter was adopted as a filter applied to estimates at the quarterly level.

8. After implementing these changes it was still found that the household estimates using the propensity method differed from census estimates (estimates for 30 June in census years based on the census in that year, as opposed to estimates based on the previous census). To account for this difference, a weighted average of the census estimate and the quarterly estimate is produced, with the weights being a function of the time since the previous census.

9. For example, for data after the 1996 Census (and whilst there is no data yet from the 2001 Census) the quarterly estimate of usually resident households is as follows:

Let, SM96=smoothed household estimate for 30 June 1996
    C96= Census household estimate at 30 June 1996
    SMHHq=quarterly smoothed household estimate post 1996
        where q=quarter required
    HHq=final household estimate

then, HHq=SM96/C96*SMHHq.

10. This method is applied to all quarters after June 1996 for each geographic and household composition combination.

11. For estimates in the 1991-1996 intercensal period, a combination of 1991 and 1996 adjustments were applied to quarters between the two censuses for each geographic and household composition combination as follows:

Let, SM91=smoothed household estimate for 30 June 1991
    C91=Census household estimate 30 June 1991
    SMHHq=quarterly smoothed household estimate 1991-1996,
      where q=quarter required
    HHq=final household estimate

then, for example for September 1991, one quarter after the 1991 Census and 19 quarters before the 1996 Census:
      HHS91=((19/20)*SM91/C91+(1/20)*SM96/C96)*SMHHS91
      and March 1996, 19 quarters after the 1991 Census and one quarter before the 1996 Census:
      HHM96=((1/20)*SM91/C91+(19/20)*SM96/C96)*SMHHM96

Details of the previous household estimates and the estimates resulting from the improved method are presented in the following table.

Results of the improved method

TABLE 1 : PREVIOUS METHOD AND IMPROVED METHOD HOUSEHOLD ESTIMATES by Part of State , as at 30 June

Part of StateMethod
1992
1993
1994
1995
1996 (a)
1997
1998
1999

Sydney Previous
1302663
1339786
1362752
1382367
1395131
1443960
1435686
1480702
Improved
1307098
1334794
1355551
1380129
1395131
1423522
1433382
1461193
Balance of NSWPrevious
822704
833539
847622
855969
873501
896016
901982
921752
Improved
816758
828470
842174
853643
873501
894280
896950
915579
MelbournePrevious
1134179
1142563
1164500
1184030
1199271
1224180
1245368
1254568
Improved
1128957
1147303
1163659
1182157
1199271
1217703
1236170
1247677
Balance of VicPrevious
461207
833539
478494
482970
477922
489666
489325
493070
Improved
464805
476812
480180
480520
477922
481661
482998
487698
BrisbanePrevious
502869
515433
528144
546472
558210
569958
597930
612959
Improved
492882
509711
527084
546416
558210
575533
601983
611634
Balance of QldPrevious
593099
620824
641027
653171
673907
691712
716942
725483
Improved
592051
617819
638902
653633
673907
690238
710792
722775
AdelaidePrevious
413964
413572
425771
434476
427632
439063
448444
453855
Improved
406468
412255
424526
429712
427632
438184
445314
449453
Balance of SAPrevious
146565
150943
149271
149536
152233
153836
157184
155914
Improved
147999
152596
150735
151124
152233
153477
156500
155900
PerthPrevious
438520
457379
466206
483019
486458
508104
515620
535499
Improved
435762
453079
463596
478508
486458
500578
510280
526541
Balance of WAPrevious
158848
159980
166756
166308
166528
173947
178334
183489
Improved
156445
158770
163571
162661
166528
172792
176007
181033
HobartPrevious
72844
73182
74784
75439
75843
77214
78570
77624
Improved
72162
73161
74246
75464
75843
77116
77896
77581
Balance of TasPrevious
103077
104811
106529
106892
106857
110105
107205
108648
Improved
102387
103661
105764
106837
106857
108851
107553
109115
NTPrevious
51207
54318
58950
59392
56798
62491
63701
64687
Improved
52143
53655
56157
56512
56798
61017
61104
62148
ACTPrevious
100261
104761
108001
110266
111824
115534
119289
117290
Improved
102482
106064
107995
110346
111824
115191
118284
118202
Australia Previous
6302007
6445945
6578807
6690307
6762115
6955786
7055580
7185540
Improved
6278399
6428150
6554140
6667662
6762115
6910143
7015213
7126529

(a) 30 June 1996 represents data based upon from the 1996 Census of Population and Housing


Outcomes of the improvement process

12. There are several noteworthy outcomes of the improvement process. Firstly, the "kink" that occurs after intercensal revision of both ERP and household estimates was removed, so that the annual series of data are smoother. Secondly, the annual variation in household estimates has been reduced, and the remaining changes are more readily explainable by observed changes in the household population. This drop in variability is particularly relevant for sub-national estimates or estimates of households by household composition.

13. However, as well as the improvements noted above, there are other findings of importance. In particular, the smoothing technique used means that the series, and particularly then end point of the series, is subject to revision in future releases. Secondly, the changes introduced so far are primarily an attempt to reduce the impact of using the monthly Labour Force Survey as a source for household composition propensities. This improvement process has demonstrated that the household estimate data is limited by the Labour Force Survey data, and is subject to the same restrictions regarding sampling error and disaggregation as employment statistics from that source. Estimates for Darwin are still highly volatile, despite the changes introduced and are therefore not available as standard ABS output.

Differences from other ABS Household estimates

14. The improved estimates to be published in the Australian Demographic Statistics, December Quarter 2000 (Cat. No. 3101.0) differ from estimates published in other ABS publications. In particular, the censal year (1996) estimates presented here differ from the censal year (1996) estimates presented in Household and Family Projections, 1996 to 2021 (Cat. No. 3236.0). This difference arises because of the difference in the methods employed. In determining the population base used for the projections data, household estimates were determined from living arrangement propensities, which then gave family propensities, and finally household propensities. Since the focus was initially on living arrangement type, there is an issue of classification of people away from the dwelling on Census night. For the improved household estimates presented here, this classification issue does not arise. Unlike living arrangement, household composition can be readily determined through the use of the "absent persons" question (Q 41) in the 1996 Census of Population and Housing. The estimates presented in Australian Demographic Statistics (Cat. No. 3101.0) are considered more accurate for 30 June 1996.

15. Estimates of the number of households are also sometimes presented in the results of ABS household surveys. These will differ slightly from the estimates presented in Australian Demographic Statistics (Cat. No. 3101.0). In some cases there is a difference because the scope of the survey differs from the usually resident household concept. In other cases the estimates will differ from those presented here because the information used in survey estimation (the "benchmark") is a short term projection, rather than an estimate as presented here. The family estimates and estimates of household population produced annually from the monthly Labour Force Survey, as presented in Labour Force Status and Other Characteristics of Families, Australia (Cat. No. 6224.0), use a different method from that presented here and as a consequence the improved estimates presented here again differ slightly. In all cases, the improved estimates presented in Australian Demographic Statistics (Cat. No. 3101.0) are considered more accurate estimates of both the number of households and the household population in Australia.

Data availability

16. The improved series of household estimates will be released in Australian Demographic Statistics, December 2000 (Cat. No. 3101.0), published on 7 June 2001. Data for Australia by household composition is presented in Table 24 and household totals by Part of State is presented in Table 25. The latter is also available electronically in AusStats spreadsheets, or via ABS@.

17. Quarterly estimates of all combinations of household composition by geographic area are available on request by contacting client.services@abs.gov.au. More information on household estimates can be obtained by contacting peter.routh@abs.gov.au or matthew.berger@abs.gov.au

Demography Program
June 2001