Jörgen Dalén Statistical Consulting Report 14(14)

June 30, 2003

Applications of computer hedonics in statistical agencies

This note tries to describe and compare the development of hedonic methods for computer equipment, mainly PCs, in a number of countries, where we are aware of such methods being applied. In this note we draw out some common features of this development as well as issues that are treated differently. Ultimately, the aim is to work towards best practices in the field of computer hedonics in official price indexes.

In at least six countries, hedonic methods have been introduced into official price indexes by now. These are:

·  USA: In the PPI for the output of the computer industry since 1991. The procedures are described in Holdway (2001) and additional information obtained in oral communication. In the CPI, the PPI hedonic model is applied for desktop PCs since 1998.

·  Canada: A price index for sales of PCs to the commercial and government sectors has been used since 1996. Barzyk and McDonald (2001) describe the developments and plans for the future. A draft documentation is presently available (Statistics Canada, 2003).

·  France: A price index has been included in the PPI for PCs since 1990. The methodology today and development plans are described in Evans (2002a and b).

·  Sweden: In the Import Price Index for PCs in 1991. The early methodology is described in Dalén (1992). There is no documentation of later developments. Statements and remarks below are based on oral communication.

·  Germany: In the CPI for PCs since June 2002. The new methodology is described in Linz (2002).

·  United Kingdom: In the PPI and the HICP since February 2003. The methodology is described in Ball and Allen (2003). A description of the developmental effort is made in Ball et al (2002).

Major developmental efforts are presently also made in:

·  The Netherlands: Preliminary analyses are made by van Mulligen (2002).

·  Australia: Intensive preparations for a PC price index are described in a paper by Poh and McKenzie (2002) et al. A decision on application is planned during 2003 but is not yet taken at the time of writing this report.

Below we will treat various aspects of the methodology and operations one by one and discuss what we see as the important differences between the various national approaches. We will refer to countries rather than to authors for the sake of easy understanding, but this implicitly means the reports stated above and in the reference list. Before doing this we describe the more exact coverage of the different countries as well as the results over a 10-12 year period for those three countries that have used a hedonic method throughout that period: the US, France and Sweden.

Product coverage

In 1991, the US Producer Price Index (PPI) expanded its coverage to include the output of the computer industry. The PPI constructs separate hedonic models for portable computers, personal desktop computers, entry-level/mid-range computer servers, and large-scale computers. The US CPI uses the PPI models for desktop computers only.

Canada has hedonic price index series covering two areas, computers and computer peripherals. For computers, two aggregate series are produced, based on the type of purchaser—government and business, and consumer. The government and business category has three sub-components (desktop computers, portable computers and servers) while the consumer category has two (desktops and portables). For computer peripherals there is one series for printers, composed of three types of printer technologies (laser, inkjet and dot matrix). Another series, for monitors, is sub-divided by technology into cathode ray tube (CRT) monitors and liquid crystal display (LCD) monitors.

The French PPI uses separate hedonic models for desktops, laptops and servers since 1996. This will be done also with the new system with IDC data.

In the UK PPI both desktops and laptops are covered. For the HICP it covers just desktop PCs but in principle will include laptops once these have been shown to have sufficient weight to enter the index.

Sweden covers both desktops and laptops in its Import Price Index, with one hedonic model only.

In Germany, only one model, presumably for desktops, is estimated.

Australia, in its experimental calculations, uses one model for desktops and one for laptops.

The experiments in the Netherlands have used separate models for desktops, laptops and servers.

Summary: Three countries have at least three separate model categories of computers – desktops + laptops + others (servers or large-scale). Four cover desktops and laptops. One country covers only desktops (this is only for the CPI).

Discussion: In a PPI, or more exactly an index for capital expenditures, the range of computer equipment is very large. The country that has attempted to cover most of this is the US by having separate strata for large-scale computers, mid-scale computers, desktops and laptops. Canada’s approach is also ambitious covering also two categories of accessories by hedonic models in addition to computers. In Table 1, weights from the US PPI[1] are provided which show that PCs+laptops take up about 50 % of the PPI weight for electronic computers but only 16 % of the weight of computers and computer equipment. Data for other countries are not available but are likely to be broadly similar.

In the CPI domain, i.e. private consumption, the dominance of PCs, and mainly desktops, is greater. In the area of information processing equipment (coicop 9.1.3) other products of some significance are laptops, printers, scanners, palmtops. Results from a questionnaire to EU Member States in November 2002 showed that PCs took up about 75 % of the weight of Coicop 9.1.3 (Information Processing Equipment).

The complications in estimating price changes for other computer equipment are not likely to be smaller than for PCs. For PPIs, and for the investment domain of National Accounts, a more diversified approach to hedonic modelling may be needed.

EHC relevance: EHC research has so far concentrated on PCs. With a CPI this priority seems to be well-grounded but the PPI and therefore also the investment domain in the National Accounts is more diversified.

Hedonic index results

In Diagram 1 PC price changes for PCs are compared in the three countries, where hedonic regression has been applied. The comparison is over the period January 1990 to December 2002.

Taking averages over the period December 1992 to August 2002, where we have data for all three countries, we obtain an average annual price decrease of 41.8 % for France, 47.4 % for Sweden[2] and 36.8 % for the USA. During the period the US dollar rate increased against both the other currencies and the French Franc (euro) against the Swedish Krona[3]. From the currency changes we would thus expect price decreases to be the largest in the US and smallest in Sweden but the results are in the reverse order.

Limiting instead the comparison to the last five years, Aug 97-Aug 02, we obtain an average annual price decrease of 38.0 % for France, 45.4 % for Sweden[4] and 41.4 % for the USA. The currency movements for this period have again been in the same directions as for the full period. Sweden’s large price decreases are therefore still unlogical in relation to its currency movement.

In Table 1, we present the relative shares of subcategories of category 115 (computers and computer equipment) in the US for 1997 which is the reference period for the 2002 PPI. We see that PCs and laptops account for only 16 % of all of category 115. This situation could be expected to be roughly similar also for European countries, at least if we consider an index for capital expenditures or domestic supply instead of a PPI.

In Diagram 2, PPI (IMPI), index series for the larger group of computers and computer equipment are presented for the US and Sweden. What is important to note here is the much smaller rate of price decreases at this level. Over the period December 1992 to August 2002 the annual average price decreases in the USA is 13.3 % and in Sweden 6.5 %. (The rates were almost identical for the five-year period of Aug 97 – Aug 02.) The relationship between Sweden and the U.S. is reversed compared with the PCs. This is explained by the fact that in the US, hedonic methods have been applied for most computer categories whereas in Sweden they were only used for PCs.

Taking Table 1 and Diagram 2 data together gives a price decrease for other computer equipment than PCs in the US of about 8 %. This sharp difference between price change for PCs and for other kinds of computer equipment should be carefully noted.

Eurostat commissioned a study of HICP practices for PCs (Haschka et al, 2002), which showed that the average weight of the coicop category 9.1.3 (information processing equipment) is 0.4. PCs cover 76 % of this weight and is thus, in contrast to the PPI situation, the dominant subcategory. Still we would expect a PC price index to fall more rapidly than an index for 9.1.3 as a whole.

Differences in annual price change for PCs in the HICP in the period 2000-2002 are from -4.6 % to -26.6 over 11 EU Member States, as reported in Haschka et al (2002). This is a considerably larger difference than that between the three countries with hedonic methods, as given above.

Discussion: The contrast between the differences between national price indexes for PCs and currency morvements strongly suggest that these differences are due to methodological effects. There is a big difference in the amount of qualified resources that have gone into the production in the three countries. In the US there has been a continual, strong involvement of qualified research staff, whereas in Sweden this involvement has been more sporadic. The French situation in this respect is not known. There is therefore a need for qualified staff, at a research level, being involved in the ongoing production of hedonic indexes. This is a strong argument for a centralised solution at the European level.

At the same time it should be noted that differences resulting from the different national approaches are at least broadly similar in comparison to the differences between various traditional (matched model) approaches, as shown in Haschka et al (2002).

The differences between price changes for PCs and the larger group of computer equipment in the US PPI shows that it is essential to clearly distinguish the different subgroups of computer equipment and make separate estimates of price change for each. For example, it is clearly unsatisfactory to let a PC index represent the whole of computer equipment, especially in a PPI.

Type of hedonic model. Imputation/time dummy

The basic choice is between a time dummy method, where the price change is a direct result of the regression equation and the imputation method, where the coefficients are used for adjusting the price differences between different models at replacements[5]. All countries now applying hedonics use an imputation price index, both in the CPI and the PPI.

Australia and the Netherlands, however, argue for the advantages of the time dummy method. As one of the advantages of this method it is claimed that no explicit use of the coefficents of characteristics is needed since the price change is estimated directly from the equation.

The time dummy method requires that price changes and hedonic coefficients are estimated from the same data set, whereas the imputation method allows the use of different data sets for the hedonic function and for the price changes, respectively.

Discussion: Van der Grient. and de Haan (2003) and van Mulligen (2002) argue that the best hedonic index is a chained adjacent period time dummy index, where also new models and characteristics could be brought in each month. However, this index requires that the estimation of the hedonic model and the index estimation is done in the same step, based on the same set of data.

EHC relevance: As long as the EHC does not want to estimate the final index, the time dummy approach is out of scope. Instead the approach has to be to prepare for estimating coefficients directly. What Triplett (2003a) calls the hedonic quality adjustment method is a straight-forward way to impute in the absence of matched models, since it is a natural extension of the traditional quality adjustment technique used since long in national statistical agencies. However, if the Dutch argument about the theoretical superiority of the chained adjacent period time dummy method is accepted, a case could be made for central estimation of the price index as well. This would set a higher requirement for the data set used for this purpose, since it also needs to be representative in all respects to the national markets.

Frequency of updating the estimates of the model

In the US, models are updated every 3-6 months[6], based on the most recent data. The time period in which the observations are recorded that are included in the regression model is as short as possible, usually only a few days.

In Canada, updates have been “occasional” prior to 1999 and twice a year thereafter. “Currently, the hedonic equations for all desktops, portables, and servers are updated quarterly and the estimates are used for the three subsequent months. The printer and monitor estimates are updated twice a year.”[7]

UK plans to update the models when necessary subject to a minimum of three times a year. The hedonic model will be updated more frequently if a monthly analysis of predicted prices against actual prices shows that the model no longer accurately predicts PC prices.

France updates the whole model quarterly, at each instance of index estimation.