Abstract

Household surveys run by the Australian Bureau of Statistics (ABS) use a sample frame drawn from an Address Register, which as its name implies is simply a list of physical residential addresses without any details about the household or its members. This means essentially surveying ‘blind’, without knowing about the occupants before contact is made to the household through physical or electronic means. Previous sample designs have though been able to draw on information at an area level, for example from the Census, to target samples towards areas with higher concentrations of particular subpopulations of interest.

As response rates decline over time, and the ABS transitions away from a reliance on face-to-face interviewing to a digital-first approach (eg eforms supplemented by telephone interviews) there is a greater need to understand how well the responding sample represents the populations of interest. In particular, we are interested in predicting and mitigating any non-response biases that may arise, in order to maximise the quality of the survey estimates.

The ABS now hosts the Person Level Integrated Data Asset (PLIDA), a secure database combining mostly administrative information at a person (and household) level on health, education, government payments, income and taxation, employment, and population demographics (including the Census) over time. This opens up new possibilities for minimising non-response biases in ABS household survey estimates.

By augmenting the sample frames with PLIDA (or a subset of it), our sample designs can take advantage of targeting households who are likely to have occupants with characteristics of interest, and we can also model probabilities of response and oversample households with lower response propensities. PLIDA-enhanced survey frames can also be used during survey estimation, to drive follow-up strategies or mode offers eg by using response propensities by mode. A third way of using PLIDA-enhanced survey frames is in estimation, to compare the responding sample to the populations of interest, and identify and adjust for biases in the responding sample.

This talk will discuss these three uses of enhanced survey frames to reduce non-response biases in more detail.