«by Brent R. Moulton* Bureau of Economic Analysis U.S. Department of Commerce Washington DC, 20230 USA June 2001 Abstract: The method of using ...»
The Expanding Role of Hedonic Methods in
the Official Statistics of the United States
Brent R. Moulton*
Bureau of Economic Analysis
U.S. Department of Commerce
Washington DC, 20230 USA
Abstract: The method of using regressions of prices on characteristics to adjust for
quality changes has grown dramatically in the United States statistical agencies in recent
years. For example, currently 18 percent of the final expenditures in gross domestic
product is deflated using price indexes that use hedonic methods. These indexes are produced by at least four statistical agencies (the Bureau of Labor Statistics, the Bureau of the Census, the Federal Reserve Board, and the Bureau of Economic Analysis). This paper details the adoption of hedonic methods by each of these agencies and discusses some misconceptions about the role of hedonic methods in estimation of price indexes.
*The views expressed in this paper are those of the author and do not represent an official position of the Bureau of Economic Analysis.
1 Introduction In May 1988, on the occasion of the 50th anniversary of the Conference on Research in Income and Wealth, Jack Triplett presented a paper on hedonic methods in statistical agency environments (see Triplett 1990). His conclusion was rather pessimistic—after more than 25 years of scholarly research on hedonic methods, which largely had resolved the theoretical and practical problems associated with the hedonic method, Triplett was only able to cite three instances in which U.S. statistical agencies had adopted hedonic methods for calculating price statistics.
The present paper revisits the role of hedonic methods in U.S. statistical agencies a decade after Triplett (1990) and reveals that the environment has changed rather dramatically. With relatively little fanfare, the use of hedonic methods has been growing at an accelerating rate. Thus, when Landefeld and Grimm (2000) mentioned that 18 percent of U.S. GDP final expenditures are now deflated using indexes that are calculated with hedonic methods, several readers treated the statistic with incredulity. This paper documents the growth in hedonic methods within the U.S. statistical system, briefly discusses some of the factors that have contributed to their adoption, and then reviews several misconceptions about hedonic methods in the context of adjusting for quality changes in constructing price indexes.
The term “hedonic methods” refers to the use in economic measurement of a “hedonic function,” h( ), (1) pi = h(ci), where p is the price of a variety (or model) i of a good and ci is a vector of characteristics associated with the variety. The hedonic function is then used in one of several ways to adjust for differences in characteristics between varieties of the good in calculating its price index. The hedonic function is usually estimated by regression analysis.
The traditional index method for controlling for quality change is known as the “matched model” method. This method controls for quality change by conditioning on the detailed characteristics of the models in the sample, ensuring that exactly the same models with the same characteristics are priced each period. Although this method conditions on the characteristics (thus requiring that the statistical agency accurately and thoroughly describe the characteristics of the model), it does not require knowing the relationship between characteristics and prices. The problem with the matched model method arises when new models appear or old models disappear; the methods traditionally used to link the samples containing the old and new models may overstate or understate the true quality differences.
Most applications of hedonic methods by U.S. statistical agencies use “composite” methods; that is, a matched model method is used to calculate the index when a full sample of prices is available for both periods, and the hedonic function is used to augment the sample for new or disappearing models when prices are not available. Exceptions are the Census Bureau’s single family house price index and BEA’s multifamily house price index, which are calculated directly from the hedonic function using characteristics as weights, the Federal Reserve’s indexes for LAN routers and switches and the hedonic portion of the BEA software index, which are calculated from the regression coefficients of indicator (or “dummy”) variables for years, and the CPI rent and homeowners’ equivalent rent indexes, which use the hedonic coefficients to adjust all units in the sample for the effects of aging.
2 A Brief History of Hedonic Methods in U.S. Statistical Agencies
The origin of hedonic methods in U.S. official price statistics goes back to the famous article by the late Zvi Griliches (1961) that was published in the report of the Price Statistics Review (Stigler) Committee. I purposely use the word “origin” because Griliches’s (1990) retrospective paper on the subject begins with a lengthy disclaimer that his work on hedonic price indexes was not particularly original; in particular, Waugh (1928), Court (1939) and Stone (1954, 1956) preceded Griliches in developing and applying hedonic techniques. Stigler (1955), however, has claimed, “Scientific originality in its important role should be measured against the knowledge of a man’s contemporaries. If he opens their eyes to new ideas or to new perspective on old ideas, he is an original economist in the scientifically important sense.” In this sense, the work of Griliches (1961) was surely original—as Lipsey (1990) observed, he took an unconventional method that was then on periphery of price statistics and demonstrated to the economics and statistics community that it could be used to address critical quality adjustment problems that previously had been considered intractable.
Following Griliches, hedonic methods quickly grew to be a new branch of economic research, which is now far too vast to be easily surveyed.1 Articles that provide overviews of some of this literature include Griliches (1971, 1990), Triplett (1975, 1987, 2000), Berndt (1983, 1991) and Bartik and Smith (1987). I will leave it to others to comment on this literature, except to note that there have been a number of theoretical and empirical controversies. As Triplett (1990) observed, however, many of these controversies have counterparts in the traditional literature on economic index numbers—issues such as aggregation across individuals in constructing a social cost-ofliving index, imperfect competition in constructing an output price index, and problems in separating demand shocks from supply shocks. My own view, which echoes Triplett’s, is that all quality adjustment methods are imperfect, but regardless of these imperfections, statistical agencies need to do a better job of quality adjustment, and so these controversies should not prevent agencies from using hedonic methods as one tool in quality adjustment. As already mentioned, many years passed before the economic literature had much impact on the U.S. official price statistics.
For example, when I recently searched for the word “hedonic” in the American Economic Association’s Econlit bibliographic CD, it located 795 articles.
2.1 The U.S. Statistical System For those readers who are unfamiliar with the U.S. federal statistical system, the statistical system follows a decentralized organizational structure. There are more than a dozen federal agencies that primarily serve a statistical function, and several dozen others that collect or disseminate statistics in support of other activities.2 The hedonic price work described in this paper has taken place at four agencies.
The primary responsibility for price statistics—the consumer, producer, export, and import price indexes—resides with the Bureau of Labor Statistics (BLS). The Bureau of Economic Analysis (BEA) produces the national, international, regional, and industry accounts, so it uses component price indexes constructed by the BLS and other agencies as deflators in calculating quantity indexes for gross domestic product (GDP) and other aggregates, as well as price indexes for these same aggregates. The Bureau of the Census conducts a variety of monthly and annual economic surveys and, at five-year intervals, an economic census; it deflates a few of these statistics, such as construction and foreign trade. Finally the Federal Reserve Board produces the monthly index of industrial production, as well as various monetary and credit statistics and the flow-of-funds accounts. Because the index of industrial production is an indicator of real economic activity, many of its current-price components need to be deflated. I will describe the adoption of hedonic methods at each of these agencies, placing them in chronological order of when the method was first introduced.
2.2 Bureau of the Census – Single-Family Houses Under Construction
The Census Bureau was the first federal agency to adopt the hedonic method. In 1968, the Census Bureau began deflating single family houses under construction using a price index based on a regression of the sales price of new single family homes against housing characteristics such as the area of floor space, number of bathrooms, regional location, and central air conditioning (Musgrave 1969; Pieper 1990). Subsequently, the Census Bureau has periodically refined the regression models and index estimation methods, including a recent switch to using the Fisher formula, but the basic character of the estimated hedonic function is largely the same as it was in 1968.
It is interesting that the primary motivation for the single-family house price index was not adjusting for changing technology and quality improvements. Rather, the index was designed to address the severe heterogeneity of new houses. Each new house tends to have unique characteristics, implying that it is difficult to use the conventional price index methods of matched models even to calculate a sales price index. Indeed, prior to 1968, construction prices were deflated in large part by cost indexes based on a weighted average of input prices—materials, wage rates, and in a few cases overhead The chief statistician of the Office of Management and Budget is responsible for coordinating statistical activities across agencies, working with the Interagency Council on Statistical Policy.
costs and profit.3 As has often been observed, using input cost indexes as a proxy for output prices is unsatisfactory because they do not allow for the direct measurement of productivity; implicitly, these indexes embody an assumption of zero productivity growth. By using a model to parameterize the differences in characteristics between houses, the hedonic method makes it possible to use sales prices from a very heterogeneous sample of houses to construct an approximately constant-quality price index.
I will also note that the number of characteristics included in the hedonic function underlying the single-family house price index is fairly small. An implication is that while the index appropriately adjusts for changes in major characteristics, such as house size or number of bathrooms, it may not be able to adjust for changes in unobserved characteristics, such as improvements or declines in the quality of materials used in construction or in construction techniques.
The Census Bureau has continued to work on construction prices and provides on-going assistance to BEA in estimating its multi-family price index, but to date the single-family index is the Census Bureau’s only hedonic index.
2.3 Bureau of Economic Analysis – National, Regional, and Industry Accounts After the Census Bureau’s price index for single family houses, nearly two decades passed before the next official foray into hedonics by a federal statistical agency.
The BEA-IBM price indexes for computer equipment and peripherals were introduced into the national income and product accounts in December 1985 (Cole et al. 1986;
Cartwright 1986; Triplett 1986). The original indexes covered five types of computing equipment—computer processors, disk drives, printers, displays (terminals), and tape drives for the period 1972-84. Subsequently, a price index for personal computers was added, and a separate index was created for computer imports. The history and present status of the indexes are documented in U.S. Department of Commerce (2000).
It is interesting to note that the problem BEA addressed in its collaborative effort with IBM was more than just obtaining an improved method of quality adjustment. Prior to 1985, BEA simply had no acceptable price index for computers, so computers had been deflated by an index that was equal to 1 for all periods.
In the early 1990’s, BLS began publishing quality-adjusted producer price indexes (PPI’s) for computers and peripheral equipment (see discussion below). As these PPI’s became available, BEA used them to extrapolate the BEA computer price indexes.
Eventually, BLS indexes were used for all of BEA’s quality-adjusted computer price indexes.
Because of these same difficulties, nonresidential structures continue to be deflated in part using cost indexes. Presently the Bureau of Labor Statistics is conducting research to determine the feasibility of calculating an output-type producer price index for nonresidential structures.
The BEA computer price indexes show rapid declines during all periods; for 1959-2000, the average rate of price change for private fixed investment in computers and peripheral equipment is –17.5 percent per year. Although skeptics have occasionally questioned the rapid price declines, the BEA index has stood the test of time. Scholarly studies have generally found similar rates of price decline (for example, see Berndt, Dulberger, and Rappaport 2000; Aizcorbe, Corrado, and Doms 2000). Several other countries now regularly use the BEA computer price indexes to deflate the computer components of computer imports and capital formation in their own national accounts.
BEA’s next hedonic index was the price index for multifamily residential structures (de Leeuw, 1993). The issues for multifamily housing were the same as those for single family housing—severe heterogeneity in the characteristics of housing units leading to the use of an inadequate proxy as a deflator. Considerable research was undertaken before an acceptable hedonic function was identified. The Census Bureau has generously assisted BEA in developing and maintaining this index.