EMPIRICAL ASSESSMENT


BACKGROUND

5.1 This section provides empirical results for the multilateral and extension methods described earlier. The intent behind the empirical analysis is to test the methods outside the theoretical realm, recognise characteristics that the methods treat differently, and compare how the methods behave at different compilation frequencies. A comparison of empirical results is important in the process of evaluating the available methods against the ABS DQF.

5.2 Empirical results are presented at both disaggregate and aggregate levels. The disaggregate level refers to price indexes below the published level that depend only on multilateral and extension methods for aggregation. These are presented to isolate the effects of different multilateral and extension methods. The disaggregate indexes are also compiled on a monthly frequency to observe how sensitive results are to a higher compilation frequency. The aggregate level refers to price indexes at the published level, which combine the disaggregate price movements of both transactions and non-transactions data respondents and aggregate using the Lowe index formula. The aggregate indexes are compiled on a quarterly frequency in an effort to produce results that are broadly comparable to the published CPI.


AGGREGATING PRICE MOVEMENTS

5.3 Traditional weighting practices vary at different levels of the CPI. At the EC level (and above), expenditure weights are currently updated every six years, primarily using Household Expenditure Survey (HES) data. Below the EC level, the expenditure weights are periodically updated to reflect households changing preferences. The compilation of elementary aggregates from individual prices uses an unweighted price index formula that considers each price observation of equal importance. The application of an unweighted formula is a practical approach due to data limitations where expenditure information at the item level is not available on a timely basis.

5.4 Transactions data provides expenditure information at the disaggregate level which opens up a range of index aggregation methods that more readily account for changes in household expenditure patterns. At the most disaggregate level, transactions data provide the ability to weight individual prices according to their level of expenditure which enhances the accuracy of outputs.

5.5 The approach adopted in this publication is to use the expenditure information available in transaction datasets across both individual items and respondents. The structure proposed in this publication is provided in Figure 5.1 for a subset of products in the All groups CPI. The proposed structure includes an additional level below the EC level that weights each respondent explicitly (by expenditure). Respondents can be weighted by using transactions data as well as other available data sources that contain information on each respondent's market share across various commodities.

5.6 The proposed structure for transactions data respondents does not contain traditional elementary aggregates, instead mapping items by respondent directly to the EC level. This approach was chosen for two reasons. Firstly, transactions data is delivered to the ABS with various respondent specific classifications which make it difficult to map to the (approximately) 1000 elementary aggregates currently in the CPI. As a compromise, mapping respondent classifications to the EC level is a suitable alternative. Secondly, with a comprehensive mapping concordance constructed at the EC level, the accuracy of the price index will be enhanced since it is accounting for substitution across a wider range of items within each respondent. Moving forward, the ABS will further investigate the suitability of aggregating transactions data directly to the EC level.

5.7 With respect to non-transactions data respondents, at the EA level and above an explicit weight will be allocated in the index structure. The EA classifications and aggregation formula for non-transactions data respondents will remain the same as is currently used in the CPI.


EMPIRICAL RESULTS AT THE DISAGGREGATE LEVEL

5.8 Empirical results at the disaggregate level are presented below for a selection of ECs. Due to the sensitivities of reporting results below the published level, all results are expressed relative to an arbitrary multilateral/extension method.

5.9 With respect to reporting the behaviour of different multilateral methods, price indexes were firstly constructed on the entire dataset for a particular respondent spanning 58 time periods (period 0 to period 57). Across the majority of ECs, the multilateral methods yielded similar results. A selection of six ECs are presented below (Figure 5.2-5.7), which standardise each multilateral method (i.e. period 0 corresponds to an index level of 100), and then express each multilateral method relative to the GEKS (e.g. TPD = TPD index less GEKS index). These figures below show the TPD, QAUV_TPD and GK typically oscillate within a two index point range of the GEKS, and the direction of this difference is typically mixed across the ECs.

5.10 While the disaggregate results showed small differences among the multilateral methods, certain commodities displayed periods with large differences. An example is presented above in Figure 5.7 for the Eggs EC where the TPD, QAUV_TPD and GK were typically within 2 index points of the GEKS, with the exception of period 47 where the results departed significantly. Using available tools that express GEKS and TPD movements as the product of contributions from individual items, the price fall of the GEKS (relative to the TPD method) in period 47 is primarily due to a small subset of products experiencing simultaneous falls in price and expenditure. In the following period, these items leave the market at historically low prices and the GEKS index adjusts back to its longer term trend. This example shows that while long-term price trends are similar, there can be significant short-term departures that are driven by the way that each multilateral method uses expenditure/quantity shares to weight products. Similar instances of the GEKS departing from other multilateral methods were observed in other ECs including: Snacks and confectionery; Poultry; and Other non-durable household products.

5.11 With respect to observing behaviour when applying different extension methods, price indexes were constructed for each multilateral method whilst varying the extension methods. Price indexes were then expressed as a difference to the index compiled using the full window of data (e.g. TPD_HS = TPD_HS index less TPD_FULL index) in order to assess the impact of using different extension methods in isolation. Since the full window index for each multilateral method is transitive it was chosen as the benchmark to compare against the different extension methods.

5.12 The results show that the choice of extension method can have a significant impact on the time series. On average, across the extension methods assessed for this publication, the HS method tended to report results closest to those of the full extension for each multilateral method. However, this was not consistent across all ECs, with each extension method performing well for specific commodities.

5.13 Figures 5.8-5.11 below provide an example for the Snacks and confectionery EC. An interesting feature is the high variability of the direct extension method compared to the full window or methods that use a rolling window (i.e. HS, MS, WS) approach. In particular, the magnitude of the difference between direct extension method and the full window results is typically largest immediately following the link month each year (e.g. periods 0, 12, 24 etc). The empirical observation of the direct extension diverging immediately after each link month, then converging toward the full extension result over the year, may be an indication that sparse data (at the beginning of each year) leads to differences in the estimated price indexes. The direct extension method used is also characterised by a shorter maximum window (13 months) compared to the rolling window approaches (25 months) in an effort to replicate the extension method described in Chessa (2016) - this is another possible factor contributing to the differences for the direct extension method.

5.14 With respect to the rolling window approaches in Figures 5.8-5.11, the indexes are identical for the first 25 periods then begin to diverge as the different splicing methods take effect. It is also apparent that all the extension methods produce a price index that drifts lower relative to results generated when using a full window. For the Snacks and confectionery EC, the HS extension method tends to sit closest to results generated using the full window of data.

5.15 In summary, the empirical evidence at the disaggregate level shows the following features:

i) The multilateral methods assessed in this publication (GEKS, TPD, QAUV_TPD, GK) follow similar price trends over the analysis period. For some commodities, the GEKS departs from the other multilateral methods in the short-term - this is driven by the role of average matched expenditure shares to weight products.

ii) The choice of extension method can significantly impact the time series. The results of this publication lend support for the use of a rolling window approach, specifically the HS, however it was observed that all extension methods tended to drift below the full window benchmark over time. The results for the direct extension method appear to be influenced by the choice of link month.

EMPIRICAL RESULTS AT THE AGGREGATE LEVEL

5.16 Empirical results at the aggregate level are presented below for a selection of groups and ECs. The benefit of presenting empirical results that include weighted price movements of both transactions and non-transactions data respondents are to produce an inflation measure that is comparable to the published CPI. The results discussed below are compiled on a quarterly frequency for each individual multilateral method. The HS extension method is used with each multilateral method due to its favourable performance in the above sub-section.

5.17 The empirical results at the aggregate level focus on three main CPI groups where transactions data contribute significantly to the Australian CPI - Food and non-alcoholic beverages, Alcohol and tobacco and Furnishings, household equipment and services (FHES). Figure 5.12 shows the multilateral price indexes rose between 23.6% and 23.9%, while the published CPI rose 22.5% over the time series at the weighted eight capitals with the largest divergences occurring in the Tobacco EC. Comparatively, all the multilateral indexes produced very similar results.

5.18 The divergences between the multilateral price indexes and CPI for the Tobacco EC becomes most noticeable following December 2013, coinciding with the Australian Government significantly increasing the excise charged on tobacco products(footnote 1) . The primary reason for the aggregate multilateral indexes to depart from the CPI is the use of expenditure weights at the EA and product level, with the multilateral methods capturing a shift in consumer preferences over time to cheaper tobacco products. Since the tobacco excise is charged on a per cigarette basis (per kilogram for loose tobacco), the multilateral methods are registering larger price rises relative to the CPI as the tax excise has a larger price effect on cheaper (per cigarette/gram) products. The example of tobacco is an interesting study where shifts in consumer preferences have actually caused a higher inflation aggregate, as opposed to traditional substitution behaviour to low inflation goods.

5.19 For the Food and non-alcoholic beverage group the multilateral price indexes rose between 3.1% and 3.3%, while the CPI rose 2.9% over the time series (see Figure 5.13). The multilateral price indexes and CPI show similar trends over time, with some short-term differences predominantly coming from volatile items (e.g. fruit). Across the cities, the relationship between the multilateral indexes and the CPI is mixed, with the CPI sitting higher or lower relative to the multilateral results for certain cities.

5.20 Despite reaching a similar price level at the end of the time series (see Figure 5.14), the Fruit EC is the most volatile commodity within the Food and non-alcoholic beverage group, with specific quarters recording significantly different price movements. Of particular interest is the September 2015 quarter, where the CPI rose 8.2% driven by stone fruits and grapes due to seasonal fluctuations in supply. In contrast, the multilateral price index methods registered falls of between 2.5% and 3.6%, with the main contributors consisting of berry products (e.g. strawberries, blueberries). This specific price movement shows the effects of using a fixed quantity index formula at the EA level in the CPI which assumes consumers purchase the same quantity of fruit each period irrespective of relative price change. On the other hand, all multilateral methods (calculated at the EC level) take advantage of the expenditure information available from transactions data and produce price indexes that account for consumers changing their expenditure patterns.

5.21 The multilateral price indexes rose between 2.6% and 3.3% for the FHES group, while the CPI rose 3.1% over the time series (see Figure 5.15). Comparing the multilateral methods, the GEKS_HS sits marginally lower than the other methods at the group level.

5.22 In summary, the empirical evidence at the aggregate level shows the following features:

i) The aggregate multilateral methods and published CPI follow similar price trends over the time series. This finding reinforces the traditional practices used by the ABS, since the published CPI is compiled using an unweighted price index formula at the elementary aggregation level.

ii) In instances where the aggregate multilateral methods and published CPI diverge (e.g. tobacco, fruit), it is mainly driven by the multilateral methods using contemporaneous information for weighting purposes that readily capture changes in consumer purchases over time.

1 From December 2013, the Australian government introduced annual 12.5% increases to the tobacco excise, as well as changing the biannual indexation to average weekly ordinary time earnings. <back