• Summary
• Downloads
• Explanatory Notes
• Related Information
• Past Releases

MICROSIMULATION AND NON-LINEAR COST MODELLING TO PREDICT EFFECTS OF INCREASING ONLINE RESPONSE ON SURVEY COST AND OPTIMAL CLUSTERING

Online collection of household survey data is becoming an increasingly important mode for official statistics. However, predicting the cost of household multistage surveys with a meaningful online response component can be challenging with traditional linear cost models, as the relationship between cost and online response levels is expected to be non-linear. A lack of available data on costs incurred under similar levels of online uptake in previous collections can also be limiting.

Microsimulation of survey enumeration can be useful for predicting possible cost outcomes and has previously been used by the ABS for predicting optimal cluster sizes. The ABS has recently explored the use of microsimulation to predict the costs of survey enumeration at a range of online response levels under a typical cross-sectional multistage household survey design. Microsimulations considered the likely distribution of mode choices by respondents nationally, and interviewers’ behaviour as they virtually enumerated their allocated samples, including decisions about approaching households, time and distance costs incurred, and the results of each approach. Predicted cost savings increased non-linearly with increasing online uptake for several cost components related to interviewers’ travel time and distance. That is, substantial cost savings did not occur until relatively high levels of online response for some cost components. However, when there was also high telephone interviewing, cost savings were more substantial at lower levels of online response.

The ABS has also recently developed non-linear cost models to predict both total cost savings and optimal cluster size, given online response rate, within block homogeneity, and ratio of costs between first and second stages. Within block homogeneity is a measure of how similar households within a cluster are to each other in key survey characteristics, e.g. households in the same neighbourhood may tend to report similar levels of income. The cost ratio is a measure of expenses incurred by an interviewer in reaching a cluster of selected houses (first stage), compared to the cost of enumerating each household once they are already in the neighbourhood (second stage). In the models, each household’s choice to respond online was treated as a random event, so the probability that an entire cluster responds online was expressed as the likelihood of online response, raised to the power of the cluster size. Predicted first stage enumeration costs were reduced by the number of whole clusters that do not require an interviewer visit, and second stage enumeration costs were reduced by the online response rate.

Cost savings estimated from these models agreed very closely with the results of the microsimulations. Optimal cluster size predictions showed some interesting features. Declustering was generally not cheapest until relatively high levels of online response (e.g. 70-90%). Where the cost ratio was lower, as in an urban area, the predicted optimal cluster size was smaller and declustering was efficient at lower levels of online response. In a cost profile more typical of a rural area, larger clusters were preferred and declustering was not cheapest until quite high levels of online response. Similarly where clusters were less homogeneous, a larger cluster size was preferred and declustering was not ideal until very high levels of online response. At up to moderate levels of online response, optimal cluster sizes were reasonably stable. However, if the within block homogeneity was low and/or the cost ratio was high, optimal cluster sizes could be somewhat larger at moderate compared to low levels of online response, before declustering became efficient.

The ABS intends to use these modelling and microsimulation methods, along with intelligence from field experience where available, to inform expectations in survey planning of the possible implications of differing levels of online response for cost and survey design.