1504.0 - Methodological News, Dec 2009  
ARCHIVED ISSUE Released at 11:30 AM (CANBERRA TIME) 04/12/2009   
   Page tools: Print Print Page Print all pages in this productPrint All

Treatment of Non-Response in ABS Business Surveys

Item and unit non-response is a problem that affects most ABS business surveys, leading to an increase in the sampling error and bias of the estimates. The two standard methods for the treatment of non-response are imputation and weighting, with the former being more commonly used for item non-response and the latter for unit non-response.

Many ABS business surveys have traditionally used a weight adjustment method equivalent to using a "live respondent mean" impute for the non-respondent units. This particular weight adjustment was implemented into ABS business surveys prior to the availability of useful auxiliary information and the introduction into estimation systems of the generalised regression (GREG) estimator, which makes efficient use of this auxiliary information. While this non-response treatment was considered the best possible solution in the situation where there was no auxiliary information available, it is inefficient if useful auxiliary information exists which can be used to reduce the sampling error and non-response bias.

An alternative two-phase calibration approach to non-response has been proposed, where the sampling mechanism is considered the first phase and the response mechanism is considered the second phase. The proposed two-phase calibration estimator was based on auxiliary information being available for all units in the population and the live/death status of a unit being known for units selected in the survey. It has been shown that if useful auxiliary information exists, then the generalised regression estimator leads to considerable sample size reductions, and the two-phase calibration approach to non-response leads to further sample size reductions.

An evaluation on the Annual Integrated Collection (AIC) demonstrated that the proposed two-phase calibration estimator can achieve reductions in the relative standard errors in the order of 12.5% and potentially reduce the non-response biases from 1.2% to 0.3% for the key data items. Alternatively, the proposed two-phase calibration estimator could be used to achieve considerable sample size savings of approximately 1,000 units, while at the same time achieve reductions in the relative standard errors in the order of 4%.

It is a relatively simple exercise to implement the proposed two-phase calibration estimator into the estimation systems used for ABS business surveys, as well as implement modifications to the existing rescaled bootstrap procedure, so that it is possible to produce approximate variances for the two-phase calibration estimator. The current estimation systems consist of a set of modular components that perform specific processes which can be easily interconnected to produce estimates from a range of estimators. The development of several simple weighting components will enable the estimation systems to produce estimates and variances for the proposed two-phase calibration estimator.

A paper on the proposed two-phase calibration estimator, recently presented to the Methodology Advisory Committee, will soon be available on the ABS website.

For more information, please contact John Preston on (07) 3222 6229 or john.preston@abs.gov.au.