1504.0 - Methodological News, Jun 2008  
ARCHIVED ISSUE Released at 11:30 AM (CANBERRA TIME) 13/06/2008   
   Page tools: Print Print Page Print all pages in this productPrint All

Sample Reductions in MPS, Boom-Time for HSM

The Australian Statistician Brian Pink recently announced budget savings for the 2008/09 financial year that affected most areas of the Australian Bureau of Statistics (ABS). One of the significant components of these savings lies in a 22 per cent reduction to the sample size in the Monthly Population Survey (MPS). The Labour Force Survey (LFS) is the most significant survey to be run off the MPS but there are also monthly supplementary surveys that will be affected by these sample cuts.

The 2006 MPS redesign, phased in from November 2007, was the first instance of significant sample reduction in the MPS for some time. The introduction of Composite Estimation (CE) for LFS in April 2007 resulted in sampling error reductions on key LFS estimates. These sampling error reductions were translated into cost savings by reducing the sample for the 2006 design and maintaining sampling error levels prior to the introduction of CE. The distribution of sample across states was also affected by the 2006 MPS redesign, with Northern Territory (NT) coming out with more sample and Western Australia (WA) with less. Unfortunately MPS supplementary surveys do not utilise CE as they are run in a single month only. The savings realised in the LFS by reducing sample and using CE simply meant increased overall sampling error in the MPS supplementary surveys.

The 22 per cent sample cuts are much less discriminating - they are effectively blanket reductions across all states and territories. However, there are no innovations in estimation to counter-balance the reduction in sample, so all estimates, including those from LFS and MPS supplementary surveys, will be subject to larger sampling errors. Recent work by Household Survey Methodology (HSM) has looked at best anticipating the impact on LFS estimates under different phase-in strategies as well as assessing operational issues associated with such a large reduction in sample size to an ongoing and historical survey.

The 2006 MPS redesign team had already engineered which Private Dwelling (PD) deselections would need to take place to implement the proposed 22 per cent sample cut. The first component of HSM's work into this sample cut was to generate Special Dwelling (SD) and Indigenous Community Framework (ICF) deselections. SD deselections were straightforward in that the set of SD selections were skipped through, with roughly 1 in 5 SD groups being deselected. For ICF, the task was not as straightforward because in states where Indigenous communities are selected, only NT had more than 2 communities initially selected. In NT, 3 from 14 communities were randomly deselected. For South Australia (SA), Queensland (Qld) and WA, where only 1 or 2 communities were selected in the first place, a strategy has been put in place for interviewers to systematically deselect 1 in 5 dwellings from the selected communities.
Some work went into assessing the immediate impact on LFS estimates resulting from different phase-in strategies. Using algebraic and empirical information, a longer phase-in period allows for greater savings and less impact on month-to-month movements under CE, but yielded a greater cumulative increase in sampling error. It was decided to use a single month "flip-in" approach as it is operationally simpler, especially with regards to interviewer management, yet is likely to yield a larger impact on movement estimates in the month of the flip-in. This strategy does, however, allow for supplementary surveys being run up to June 2008 to be unaffected by the 22 per cent sample reduction.

Up until now, the LFS has not been susceptible to small cells (those cells with small sample counts) during weighting due to its large overall sample size. With the 2006 sample redesign and subsequent 22 per cent sample reduction, an assessment was undertaken on how likely the LFS was to encounter benchmark cells with small or zero sample count. Binomial probabilities were used to simulate how often this may happen under the reduced sample and as a result some benchmark cells have been recommended for collapsing in LFS, as the LFS processing team does not have sufficient time every month to remedy small and zero cells on a case-by-case basis.

For more information about reduced sample in the MPS, please contact Justin Lokhorst on (08) 8237 7476.