Methodological News, June Quarter 2023

This issue contains two articles:

• Knowledge graphs for consumer dataset exploration
• An alternative analytical approach to trend estimation - accounting for the COVID-19 impacts to Labour Force Survey time series 

Consumer data is a not a traditional data source for national statistical offices. This type of big data can be rich in detail but by its very nature comes with a series of challenges. The potential use of transactions data to derive representative consumer spending insights has recently been investigated.

Two approaches were taken in exploring the potential value of these types of complex datasets for statistical purposes. The first was to focus on the most robust elements of the data, to see what sorts of aggregate inferences could be derived from them.  The other was to examine the structure of the data and the many-to-many relationships between units (such as those among accounts, merchants, account demographics and expense demographics), to understand the limitations of these data and inform thinking around their potential future use.

An advanced information system developed at the ABS for integrating and analysing diverse multisource data was used to enable the second of these investigations. In this case, it facilitated the representation and manipulation of the consumer data while combining these with relevant statistical classifications in the form of a network construct called a knowledge graph.   As well as containing all the source and derived data items together, the knowledge graph contains the underlying concepts of the data and associated metadata. Exploration of the data through a knowledge graph construct enabled the data to be connected and visualised at a very detailed level.

Through this approach, real world complexities associated with consumer spending and income could be analysed.  These complexities include accounts that are shared within and across households, the implications of connected transactions, and infrequent transactions (such as education expenses and large one-off purchases) which may or may not be present in the data capture period.  Special cases where a particular transaction type can represent either an expense or an income – for example, child support, gambling and insurance - could also be interpreted more readily.

The ability to load these data into a knowledge graph also provided the means to iteratively inform the potential reuse and redesign of income and expenditure data items. Additional insights were provided by taking a fortnightly view of data, giving new insights on the seasonality of transactions and the impact of other factors like lockdown periods during COVID.  Importantly, where reality differed from expectations, this was able to be better reflected in the graph structure.

As the ABS continues to explore a range of alternative data sources into the future, ABS knowledge graph capability will continue to be developed as a key tool to draw together different types of data and to derive more flexible insights from complex integrated data assets.

For more information, please contact Bernadette Giuffrida.

Trend estimation is a crucial tool for understanding the underlying long-term movements and identifying turning points in a time series. The unprecedented impacts of the COVID-19 pandemic have presented challenges for producing reliable trend estimates of many economic time series.

A time series can be decomposed into three components. The trend component is a measure of the medium to long-term direction of a time series, in contrast to the seasonal component (systematic and calendar related movements) and the irregular component (unsystematic and short-term fluctuations). As part of estimating the trend, significant events are assigned to either the trend, seasonal or irregular component. At the time, it was not known whether the impacts of COVID-19 would be short or medium to long term, therefore, could not confidently be assigned to the appropriate component.

Due to this uncertainty, from March 2020 to September 2022 the Australian Bureau of Statistics (ABS) suspended publication of Labour Force Survey trend estimates and explored alternative methods for the estimation of the trend. One such effort was to investigate the use of a Bayesian structural time series model. This new method has several advantages over the current X12-RegARIMA method, particularly its ability to automatically handle COVID-19 impacts by identifying outliers in each of the basic structural time series components. The new model also allows time-varying variance, and dynamically determines the responsiveness and robustness of the trend component.

The method was evaluated on a selection of ABS Labour Force Survey time series. The case studies demonstrated that the Bayesian structural time series model was a useful tool in guiding the treatment of extreme shocks from the COVID-19 pandemic.  The model’s confidence intervals were used to prioritise series with the most significant differences between the traditional trend estimate and the alternative trend estimate for analysis. Together with existing knowledge about the series, the alternative trends were used as extra information in decision making when deciding how to handle shocks to the Labour Force data.

The Bayesian structural time series model was not intended to replace the existing X12-RegARIMA seasonal adjustment and trend estimation method. The model has instead been used to facilitate annual seasonal adjustment reviews, helped to explain pandemic-related trends in a more coherent manner and has the potential to effectively handle future extreme shocks.

For more information, please contact Linh Huynh.

Please email methodology@abs.gov.au to:

• contact authors for further information
• provide comments or feedback
• be added to or removed from our electronic mailing list

Alternatively, you can post to:

Methodological News Editor
Methodology Division
Australian Bureau of Statistics
Locked Bag No. 10
Belconnen ACT 2617

The ABS Privacy Policy outlines how the ABS will handle any personal information that you provide to us.

Releases from June 2021 onwards can be accessed under research.

Releases up to March 2021 can be accessed under past releases.