National Academies Press: OpenBook

At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes (2002)

Chapter: 4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change

« Previous: 3 Index Domain
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 106
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 107
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 108
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 109
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 110
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 111
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 112
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 113
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 114
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 115
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 116
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 117
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 118
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 119
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 120
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 121
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 122
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 123
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 124
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 125
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 126
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 127
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 128
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 129
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 130
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 131
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 132
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 133
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 134
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 135
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 136
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 137
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 138
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 139
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 140
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 141
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 142
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 143
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 144
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 145
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 146
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 147
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 148
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 149
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 150
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 151
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 152
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 153
Suggested Citation:"4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change." National Research Council. 2002. At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes. Washington, DC: The National Academies Press. doi: 10.17226/10131.
×
Page 154

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

4 Evolving Market Baskets: Adjusting Indexes to Account for Quality Change T he ever-changing mix and quality of products and services available in the market create difficult problems for price index construction. At a basic conceptual level, the problem is easy to understand: under either of the conceptual frameworks we have discussed, unadjusted price comparisons be- tween an item and a non-identical replacement cannot generally be treated as equivalent to comparisons that involve an unchanged item. However, developing solutions and assessing techniques for correcting the problem is extremely com- plicated. Quality change has typically been considered the least tractable problem associated with the Consumer Price Index (CPI). The pervasiveness of item replacement alone makes quality change impos- sible to ignore. Item replacement refers to the process whereby a Bureau of Labor Statistics (BLS) field agent must select and price a different product because the one previously included in the sample can no longer be found on the store shelf. Moulton and Moses (1997:323) estimate that, based on 1995 data, about 4 per- cent of price quotations on average every month involve a replacement item. Some items are replaced more than once during a year, and this translates into an annual replacement rate of about 30 percent for CPI items scheduled to remain in the sample. Although a price adjustment is not made in each case, a judgment about quality change is. In about a third of these cases—roughly 10 percent of all CPI items each year—a quality adjustment is deemed necessary. Moulton and Moses also show that, relative to continuously priced items, replacement items have a disproportionately large effect on the rate of change in the CPI. “Quality change” can take many forms. In the research literature, the distinc- tion is often made between quality change and new goods. Unfortunately, this 106

EVOLVING MARKET BASKETS 107 distinction does not create clear-cut categories that imply specific corrective approaches. For instance, is a cell phone an improved wired phone or an entirely new product? What about a high-definition television, a fuel cell automobile, or on-line stock trading? The line between “new” and “improved” is inevitably arbitrary. The situation is brought into focus somewhat by thinking in slightly different terms, framed by consideration of how CPI product sampling and item identification actually work in real cases. Following Armknecht et al. (1997), three distinct cases can be delineated: (1) A new item replaces another that has been or soon will be discontinued and that will fall out of the CPI sample. Replacement goods may be substantively similar (in which case there may be no quality issue at all), or they may be improved (or possibly inferior) versions of the discontinued item. These goods replace old goods but fall into familiar CPI categories—e.g., 2001 Fords. (2) A new “supplemental” good appears that does not replace a specific outgoing good in the CPI, but that does fit appropriately into an existing item strata category—e.g., Honey Nut Cheerios. (3) A genuinely new item appears that does not fit into an established CPI item or strata—e.g., VCRs or wireless phones. In some sense, all of the above situations involve new goods; however, the extent of the difference between an old and a new product ranges from close to zero, to run-of-the-mill quality changes that happen on a daily basis, all the way to the appearance of radically new products that reflect what Nordhaus (1998) calls “tectonic shifts in technology.” Over time, BLS confronts situations on all points of this quality change spectrum. On the easy end, a commodity analyst may be forced to compare a 2- pound bag of rice with a 1-pound bag. Perhaps the previous 1-pound bag is out of stock or is not sold much anymore. Something like this happened with butter, which used to be sold in half-pound packets and now is more frequently sold in 1- pound packets. Most economists would simply work with per-pound prices in both cases. Of course, a 1-pound package is not identical to two half-pound packages, since the former requires longer storage, may be more likely to go bad, or may be sold with size discounts, and so on. But in many cases of this sort, per- unit prices seem likely to provide a very good approximation. The BLS appar- ently agrees. For instance, when the CPI went from pricing 16-ounce cans of tomato sauce to 14.5-ounce cans, all of the difference in price per ounce was attributed to pure price change (Kokoski et al., 2000:2).1 1This is not to say that the nonlinear pricing issue is unimportant, particularly for large differences in package sizes. As a first step toward estimating its impact on the CPI, BLS could, in a straightfor- ward manner, perform empirical research that examines how unit prices vary with package sizes. Of course, this only applies to products for which a range of sizes is typical.

108 AT WHAT PRICE? Some of the distinctions between cases can be clarified by casting them in this repackaging framework (this idea dates back at least to Fisher and Shell, 1968, 1972). This framework deals with situations in which the amount of “good” in the good has changed. It is difficult to think of actual examples other than changes in package size that correspond exactly to this framework, but imagine that gasoline has been improved so that it gives a 25 percent increase in miles per gallon for all vehicles and is otherwise unchanged. Once again, the solution seems fairly clear: the real price has fallen by 20 percent from, say, 5 cents to 4 cents a mile. One useful way of thinking about this is that the good is not gasoline but miles from fuel, and the price of the latter has fallen by 20 percent. Another example might be a new razor that yields more shaves before becoming dull. These cases converge with the butter case when one shifts from thinking of gasoline or razors to thinking of a good that more directly relates to consumer welfare. Once the good is defined appropriately—which is not trivial—and one thinks of the market good as a repackaged real good, the right way to handle quality change becomes transparent. The basic idea applies to more complicated cases, though the practicalities get harder. In most cases, there is no single obvious quantitative metric (like miles per gallon or number of shaves) to use in redefining the package, which makes it difficult to identify a simple one-to-one relationship between the real goods and the market goods. Economists and marketing specialists often think of situations of this sort in terms of characteristics, with market goods consisting of various combinations (packages) of several characteristics. Since one often does not really know the characteristics—because each good has many and because there is often no nonarbitrary way of defining them—things are rarely as simple as in the gasoline case, let alone the butter case. It is conceptually useful though to think of approaches such as hedonic techniques, which we discuss in detail below, as an attempt to redefine goods so that, by repackaging, one can factor out quality change. The really hard cases occur when a new good introduces new characteristics, in which case the repackaging idea cannot help with measuring quality change. But it is unclear that any practical technique can help in these cases or, indeed, whether radically different goods can even be appropriately discussed in the context of price measurement. For instance, in no clear sense did the introduction of cellular telephones reduce the general price level. Yet that new product did increase the well-being achievable by a subset of the population for a fixed money outlay and, in that sense, reduced the cost-of-living. Our coverage of the quality change/new goods problem follows the tax- onomy outlined by Armknecht et al. (1997). First, we contrast the nature of the problem as it arises in the COGI and COLI contexts. In the next three sections, we sort though the gradations of quality change that occur along the repackaging spectrum. This discussion includes a brief review of the evidence of CPI bias presented by the Boskin commission (Boskin et al., 1996) as well as a discussion

EVOLVING MARKET BASKETS 109 of BLS item replacement procedures and their associated biases. Chapter 5 con- siders separately the case when goods appear that do not fall into existing product categories. The second half of this chapter assesses the role of hedonic regression techniques in quality adjustment. We offer specific recommendations about the applicability of hedonics for adjusting observed prices or for directly construct- ing indexes and about approaches to selecting items for quality adjustment.2 COLI AND COGI VIEWS OF THE QUALITY CHANGE PROBLEM The general problem of changing quality can be illustrated by simple ex- ample. Consider a price index for automobiles for which, in the reference period, the dominant type of automobile has a steel dashboard and no seat belts and is a gas guzzler. Now suppose that, in the comparison period, the dominant type of automobile has leather appointments, airbags, and efficient fuel economy. Direct comparison of the nominal prices of these cars will yield little meaningful infor- mation. What does it tell us if the price of a 2002 Camry is 10 times that of a 1965 Rambler? Similarly, if this year’s computer model costs the same as last year’s but does more and does it faster, what does the observed price constancy really tell us? Nordhaus (1998:59-60) points out that a fundamental problem associated with quality change is raised by these types of comparisons because “conven- tional price indexes measure the prices of commodities that consumers buy rather than the cost of attaining a given level of economic well-being or utility.” The manner in which quality change and new goods problems arise depends to some extent on the index’s underlying conceptual structure whether a cost-of-goods index (COGI) or a cost-of-living index (COLI) though procedures for dealing with these problems are essentially the same in both cases. COLI The COLI requires that prices, or the index itself, be adjusted to account for effects on living standards that accompany changes in the quality of goods and services. For certain commodities, the quantitative adjustment could be straight- forward—e.g., the new automobile fuel that increases miles traveled per gallon. But in most cases, the relationship between product or product characteristics (inputs to well-being) and actual well-being created cannot be directly observed. 2We bypass the issue of quality change as it affects nonmarket inputs to consumer well-being (things like air quality, traffic congestion, and sense of personal safety) that are not captured in conventional price indexes (see Chapter 2). In addition, while we acknowledge the theoretical valid- ity of the Boskin commission’s observation that changes in the variety of available goods and ser- vices affect consumer well-being, we know of no useful way to deal with this issue in index con- struction.

110 AT WHAT PRICE? It may be impossible to measure the value, even for just one consumer, created by the change from black and white to color television, by an increase in the user friendliness of computers, or by the addition of antismog devices to automobiles. Even setting aside the problem that the value attached to changing products may differ widely among consumers, changes in the mix of items sold raise two difficult issues. First, when outgoing items are replaced, COLI calculation re- quires isolating a pure price component from the observed price difference be- tween the outgoing item and its replacement, which reflects both pure price change and quality change. If the underlying index methodology is unable to disentangle the quality-driven price movement from the “pure” price movement, living standards cannot be held constant. Second, techniques must keep the com- posite of index items—which is in constant churn—relevant to consumers’ mate- rial well-being. The addition of new goods into the marketplace generally raises (and the elimination of goods lowers) the welfare of some consumers; until the new good is represented, this welfare change is not reflected in the price index.3 The cost-of-living approach provides a theoretical framework for thinking about problems associated with the changing nature of goods and services avail- able in the market. If a rational consumer buys two varieties of some product— apples, for instance—in some (relatively short) period, economic theory asserts that the ratio of their prices measures their relative qualities, at least at the mar- gin.4 The next logical step is to assume that such price ratios provide meaningful measures of relative qualities even if there are many consumers and some do not purchase both products simultaneously. This assumption may be misleading when notions of quality differ across consumers, since demographic changes may then shift relative prices without quality change. Without this assumption, however, there is no way to use market data to recognize that, for instance, replacing a low- price variety with a high-price variety can make all consumers better off if the new variety is of sufficiently higher quality. More generally, if the quality of goods improves on balance over time, a cost-of-living index will discount some of the nominal price increases that occur, and the overall price index will rise more slowly than the average of the unadjusted prices. 3Hausman (1997) has argued that the CPI is also biased as a cost-of-living estimator because, to the extent that consumers value variety, it makes no allowance for increases in the number of choices within index categories. Conceptually, this assertion is hard to dispute—if for no other reason than greater variety permits better matching to individual tastes, which gives some people pleasure di- rectly. On the other hand, the existence of greater variety may, in some cases, be welfare decreasing if it creates increased search costs. There is no known way to capture such effects accurately in regular index production. 4In practice, of course, it is often a matter of judgment as to whether one is dealing with two varieties of the same product or two different products that happen to be relatively close substitutes. Also, as the number of varieties multiplies, the act of choosing itself may require more time and effort.

EVOLVING MARKET BASKETS 111 Unfortunately, quality adjustment techniques seem destined always to have an ad hoc element. The aspects or characteristics of goods that determine con- sumers’ perceptions of quality are not consistently observable. Moreover, the quality and taste components of price change are often inexorably intertwined. On the other hand, BLS has to balance on a case-by-case basis the errors that inevitably arise in the adjustment process against the errors that would inevitably arise either from ignoring quality changes or from assuming, as is often done, that all price differences between similar items reflect quality differences. We return to these issues below. COGI In a conventional Laspeyres index, the changing set of available goods and good characteristics also creates problems. Once an index item from the refer- ence period is replaced by a different item, a strictly defined Laspeyres index cannot be calculated, since an identical bundle can no longer be priced in the comparison period. Given the pace at which new goods are introduced in a modern economy—ranging from those with slightly modified characteristics to those that are entirely new—it would be highly restrictive to monitor price infla- tion solely from a bundle chosen for stability. A “Big Mac” index may lead to misleading conclusions about general price movements, particularly since stag- nant and dynamic sectors of the economy are likely to display systematically different price trends. Nonetheless, in no small part because of the uses to which they are put, it seems desirable to adjust price indexes to account for changing item quality and to reflect the changing mix of goods over time. In practice, one need not be methodologically boxed in by this narrow text- book view of a Laspeyres index. And the CPI has in fact been modified—since at least 1967 when BLS began adjusting automobile prices—to address quality issues while, at the same time, maintaining its basic COGI structure. A working definition requires only that a set of market goods and services that are valued by consumers be identified for inclusion in the index. Since purchasing patterns and the set of available products have changed, the basket has been allowed to change over time. The organizing principle is the desire to cover the goods on which people spend most of their money and then to make adjustments to account for quality change. A COGI proponent is likely to argue that quality adjustment is necessary because, when the nature of goods change, prices of like items cannot be compared over time since the original bundle of goods no longer exists. A COLI proponent is likely to add that, since improved products generate higher levels of consumer satisfaction, observed prices must be adjusted to isolate changes in the cost of maintaining living standards. These differences would not affect their evaluations of alternative adjustment mechanisms. Again, the idea of repackaging helps draw some distinctions between a COGI and a COLI in handling quality change. When two half-pounds of butter are

112 AT WHAT PRICE? replaced by a 1-pound pack, a strictly defined Laspeyres index is an impossibil- ity. However, it is possible to go forward in this framework if the item for pricing has been defined as “butter” instead of “half-pound packets of butter.” In our earlier gasoline example, a Laspeyrian working in this fashion would price “miles from fuel” and not “gasoline.” The Laspeyres approach has no difficulty pricing characteristics, provided of course that one has some way of identifying and measuring the relevant ones, a difficulty that is common to all approaches. Any Laspeyres-type approach must, however, begin with a definition of the goods (which may be characteristics) to be priced. Selection must be based on some clear notion (e.g., market share) that the set of goods represents that which people buy and what gives them utility. This is as true in a world of fixed quality as in one with changing quality. It is important not to confuse the issue of the defini- tion of goods with the issue of a COLI versus a COGI. The arguments and recommendations in this chapter reflect the panel’s view that the CPI should be adjusted, for most categories of goods and services, to account for changing quality. In the next two sections, we review evidence on quality change bias. First, we briefly examine the Boskin commission report (Boskin et al., 1996) which focused on factors that are external to the CPI sample. We then review CPI methods for adjusting quality-changed items within its sample as well as potential biases associated with those methods, reserving the special case of hedonic adjustment methods for the following section. EVIDENCE FROM THE BOSKIN COMMISSION REPORT In accordance with its congressional charge, the Boskin commission ven- tured to estimate, by source and by item strata, biases in the U.S. CPI, relative to a hypothetical cost-of-living index. The commission’s report (Boskin et al., 1996) estimates quality change and new product bias (which they treat interchangeably) to be 0.612 percentage points per year—the largest component of its overall CPI bias estimate of 1.1 percentage points.5 The commission’s report has received extensive attention in the academic literature; numerous studies (both pre- and post-Boskin) corroborate the general view that quality change bias exists, though there is much debate on the size and sources of the biases. Much of the research has focused on specific index items (e.g., Berndt et al. [1996] on prescription drugs, Cutler et al. [1996] on hospital and physician services, Hausman [1997] on new cereal varieties). Shapiro and Wilcox (1996) did estimate an overall CPI bias, in the range of 0.6 to 1.5 percentage points per year, but it is extrapolated from trends for a limited number of products and not from an evaluation ranging across all CPI item categories. Unfortunately, research on the potential magni- 5Though the Boskin commission does not attempt to identify separate quality change and new goods bias estimates, the report does make some descriptive distinctions between the two categories.

EVOLVING MARKET BASKETS 113 tude of the quality change or new goods problem has not revealed broadly appli- cable techniques for correcting these biases. In contrast, a set of generally accepted methods has emerged for addressing other perceived index problem areas, most notably substitution bias. Shapiro and Wilcox (1996) describe solutions to the substitution component of the bias prob- lem as harvesting the “low-hanging fruit” of the CPI bias problem. Sticking with the harvesting metaphor, solutions to quality change and new goods bias prob- lems must be the fruit at the top of the tree, the kind that requires expensive tools to reach or that may not be reachable at all. The theoretical COLI perspective provides a rationale for tracking the value to consumers of new models and commodities and suggests why, for certain purposes, an index should be adjusted to reflect these changes. However, the COLI theory is less illuminating when it comes to directing research toward appropriate corrective techniques. Indeed, finding approaches for accurately deal- ing with changing goods and new goods is the most difficult obstacle to fulfilling the Boskin commission’s prescriptions for BLS to establish a cost-of-living index as its objective in measuring consumer prices. Reflecting the difficulty of the issue, the Boskin commission report did not advance any formal recommenda- tions about how BLS could improve its measurement of quality change.6 The Boskin commission suggested perhaps that BLS should be doing more of the things it already does to correct for quality change bias, but seemed to concede that it did not have new ideas for approaching the problem. Summarizing the commission’s report, Gordon and Griliches (1997:84) write: The difficult questions posed by quality change and the continual arrival of new products . . . have been called the “house-to-house combat of price measure- ment.” Because the magnitude of quality-change bias differs so much across product categories, any overall evaluation must be conducted “down in the trenches,” taking individual categories of consumer expenditure, assessing quality-change bias for each category, and then aggregating using appropriate weights.” The Conference Board (1999:21) study group concurred: “In an advanced, dynamic economy like ours, there is no alternative to thorough, detailed analyses that slog through the data category by category, item by item. This is difficult, costly work, but no shortcuts are available.” Such conclusions reinforce the premise that general solutions, equivalent to the use of superlative indexes or geomeans to address substitution bias, do not exist to correct for quality change 6In contrast, 5 of the commission’s 17 recommendations deal directly with a form of substitution bias—for which concrete options (superlative and superlative approximation indexes) exist. Indi- vidual commission members have elsewhere advocated expanding the use of hedonic regression methods to control for quality change for specific product types (see, for example, Boskin et al., 1998:14).

114 AT WHAT PRICE? and new goods bias. Furthermore, there is no guarantee that even such “detailed analyses” will produce results that are suitable for inclusion in the CPI. While the Boskin commission offered no new remedies, it had much to say regarding the magnitude of quality change and new goods effects on the CPI, producing a comprehensive, categorical point estimate. Of the 27 CPI item cat- egories evaluated for quality change by the commission, 8 were assigned a bias of zero; the other 19 were estimated to impart a positive bias on the index.7 Esti- mates for 6 of the 19 positive bias categories were calculated using a combination of results from existing studies of specific items and inferred figures for similar unresearched items in the category. Two upper-level CPI categories assessed in this way—appliances (including electronic) and medical services—contributed more than half the estimated overall quality bias. The commission performed original research or detailed back-of-the-envelope calculations for 4 categories. For the remaining 9 categories, empirical evidence was unavailable, and a de- scriptive approach discussing possible bias sources coupled with guesswork had to suffice (Moulton and Moses, 1997:310). (See “Technical Note 1” to this chap- ter for a review of upper-level item categories that the commission identified as contributing significantly to its overall CPI bias estimate.) BLS APPROACHES TO QUALITY CHANGE In constructing its CPI, the BLS has implemented a number of techniques to minimize perceived biases associated with its modified Laspeyres approach. For many decades—starting long before the comparatively recent calls for a cost-of- living index—BLS has been aware of problems posed by items whose quality is changing over time. In general, the agency has appealed to the cost-of-living theory in describing its efforts to confront the issue. BLS readily acknowledges that, relative to some ideal COLI, introduction of new goods and quality change of existing ones may bias the CPI in two different ways.8 First, there are biases associated with quality changes that are detected in the CPI sample and for which BLS attempts to correct. In this case the question is: “What is the bias, if any, of CPI procedures for handling quality change when quality changes appear on CPI items?” (Triplett, 1997:24). Second, there are 7See Boskin et al. (1996) or Gordon and Griliches (1997) for the complete list of estimated bias by category; see Moulton and Moses (1997) for a detailed critique of the estimates. 8As noted above, the distinction between a “new good” and a new variety or improved-quality good is arbitrary. In terms of CPI construction, we think of a “new good” as one that would require creation of a new item strata (or entry-level item) and that can only enter the index by initiation of a new item classification structure—the VCR is an example. Quality change refers to evolving charac- teristics of a good or service already included in the index and whose price can be adjusted to reflect the change at any point—a laptop computer with more memory is an example.

EVOLVING MARKET BASKETS 115 factors that go unrecognized with current CPI methods that bias the index as a representation of changes in the price to consumers of attaining a given level of well-being. A COLI in its purest sense would respond even to changes in non- market goods (such as air quality or commuting time). Moreover, even in a conditional COLI, changes occur in the market—most visibly the appearance of new goods and services—that affect well-being but are not accounted for, at least not immediately, in the CPI. Estimating the extent of the first source of bias requires evaluating internal CPI quality adjustment practices. BLS uses a range of quality adjustment ap- proaches when a new item replaces an old one in the sample; the result may add all, some, or none of an observed price change to the index. Some of these approaches implicitly adjust for quality differences; others produce direct cost- based or hedonically derived comparisons of quality that are used to adjust ob- served prices explicitly. The Boskin commission report emphasized the second sort of quality-related biases, those created beyond the CPI sample and outside of CPI methodology. They focused on one subcomponent: underrepresentation of new market goods in the CPI. One way of thinking about new goods in the context of a price index (due to Hicks, 1940) is to imagine that the good was always available but at such a high price that no one would buy it. When the good is introduced, one can calculate the effect on the cost of living by translating the new availability into a price reduction, from the (imaginary) price that choked off demand to the new (lower) price at which it was first sold. The CPI as calculated makes no attempt to capture this “price reduction” associated with the introduction of new goods (see Hausman, 1997). Nor does it attempt to capture the later similar “effective” price reductions that occur as more and more consumers learn about new goods and experience a reduction in the cost of living because of that knowledge. Since the CPI market basket has historically only been revised every 10 years or so, new goods often entered the basket only after a long delay, and early stages of product price cycles were missed. Other sources of index bias may go undetected, such as those associated with gradual change in the quality of services (medicine, educa- tion, airline travel) or intangible aspects of quality change, such as improved stereo sound or television picture quality. Estimates of the magnitude of quality change bias seem to be closely tied to the type of bias researchers emphasize. Triplett (1997) argues that the Boskin commission arrived at a high-end estimate of quality bias partly because it fo- cused primarily on biases generated by new goods (such as VCRs and mobile phones) during the periods when they were outside the CPI sample. He further suggests that current BLS methods for within-sample adjustment—which occur when an old product disappears from a CPI outlet and is replaced by a new noncomparable one—may impart some downward biases (Triplett, 1997:24): “The implications of the methods used in the CPI for handling quality changes are not well understood by economists; the CPI [Boskin] Commission did not

116 AT WHAT PRICE? discuss them adequately, and some of these methods over-adjust for quality change, so that improving quality can generate downward bias in the CPI.” Essentially, BLS methods for adjusting observed prices of items that have undergone significant quality change, as judged by a commodity analyst, borrow information about price changes observed for similar items. For example, say a new improved microwave oven replaces the old model at a CPI outlet. Under one method, BLS will assume that the pure price portion of an observed price change between the old and new models is the same, in percentage terms, as that for other microwave ovens. Any remaining price difference is attributed to quality change. Such a method would implicitly overstate the effect of quality (and impart a downward bias on the CPI) if, for instance, manufacturers tend to increase prices (beyond those that cover costs of implementing improvements) when they roll out new models. We lay out BLS quality adjustment methods and examine poten- tial biases in greater detail below. CPI Item Replacement Methods As noted at the beginning of this chapter, the manner in which goods (and services) appear and disappear can take a number of forms: old models are replaced by new ones that are nearly identical; new models are introduced that embody clear improvements over their predecessors; models may display quali- tative change in existing features or may introduce altogether new features. To accommodate some of these differences and to overcome data and procedural limitations, BLS employs alternative methods, shown in Table 4-1, for treating replacement item price quotes. For cases in which a sample item is replaced, the observed price change must be (1) considered a pure price change (e.g., simple repackaging), (2) attributed entirely to quality differences, or (3) attributed partly to price change and partly to quality change (Kokoski, 1993). Cases 2 and 3 require adjusting observed prices prior to inclusion in the index; all three require judgments by BLS com- modity analysts. Case 1 results in what BLS calls “direct comparison,” which applies when the replaced and replacement items are determined to be comparable by the commodity analyst. A repackaged food item or a new color of shirt are examples. Direct comparison is essentially item replacement for cases in which adjustment to the observed price has been deemed unnecessary. As Table 4-1 indicates, this is the most common finding. According to Moulton and Moses (1997), for 1995 about 65 percent of item replacements were in this category. With direct comparison, a commodity analyst has determined that it is appro- priate to treat the observed price change as pure price change. If any quality change does occur, its effect on the index is not filtered out. The Boskin commis- sion wrote that direct comparison, which it called “comparable substitution,” likely imparts an upward bias to the index since “in practice most goods tend to

EVOLVING MARKET BASKETS 117 TABLE 4-1 CPI Item Replacement Methods and Use Rates, 1995 Percentage of Percentage CPI Price Quotes for Change Attributable to Item Categories Replacement Items Replacement Items, for Which Method Using Method Decomposed by Methoda Frequently Applied Direct comparisonb 65 60 All categories Overlap pricing 1 10 Apparel, medical care Deletionc 15 2 Medical care, food and beverages Class-mean 8 18 Apparel, transportation Direct quality adjustmentd 11 19 Transportation, apparel, computers aThe percentage change in CPI item categories affected by item replacement was 2.16 in 1995. Of this, 1.09 percent was attributable to replacement items (leaving 1.07 attributable to continuously priced items). Thus, for instance, 0.6 * 1.09 gives the overall percent change in the CPI that could be attributed to direct comparison price quotes in 1995. bUsed for comparable replacements; the rest of the methods listed are applied to noncomparable replacements. cThis follows Triplett’s terminology; it is typically called the link method in BLS literature. dThis category includes both cost-based and hedonic methods. SOURCE: Data from Moulton and Moses (1997). undergo steady improvement, and often a better model is introduced with no change in price, causing the quality change to be missed entirely” (Boskin et al., 1996:19).9 Empirical evidence presented by Moulton and Moses (1997) implies that any upward bias from ignoring quality change under the comparable substi- tution method is small, tempering the Boskin commission view. Triplett (1997:26) synthesized the empirical evidence: “Note . . . that the average price change [shown using 1995 CPI data] for the direct comparison cases (2.51 percent) is not higher than the quality-adjusted price changes for CPI cases where a direct qual- ity adjustment is made (2.66 percent—Table 7). This suggests that the upward bias from ignoring quality in the direct comparison cases is small.” Triplett goes on to explain why this is likely to be the case: “Direct comparison is the sanc- tioned [here used to mean approved] method for cases where the quality differ- ence between varieties a and b is small, so it is reasonable that the quality errors are also small (though they might be pervasive).” 9The Boskin commission is really criticizing the BLS method for assessing comparability. Under the comparable replacement procedure, the new price is recorded, and the bias will be the same (in absolute terms) whether or not price a change has occurred.

118 AT WHAT PRICE? Neither side would disagree with the proposition that, to the extent that undetected quality changes are more often improvements than deterioration, di- rect “comparable replacement” will bias the CPI upward. However, for the bias to be large, undetected quality change would have to be distributed such that, within the sample, missed quality improvements were either much more perva- sive or of much greater magnitude than incidents of missed quality deterioration. This has never been shown empirically, possible because judgments by commod- ity analysts about comparability are not random. A new model (of, say, a micro- wave oven) with major changes in characteristics is more likely to be judged noncomparable than is a new shirt that displays only minor changes, such as in styling or color, over its predecessor. For cases deemed by a BLS commodity analyst to require noncomparable substitution—that is, when there is a perceived quality change—BLS uses one of four options (other than hedonics, discussed separately below): • overlap pricing, • explicit cost-based price adjustment, • a deletion link, or • a class-mean link. The overlap pricing option can be used when both old and new models are available in at least one period. If the new version is introduced in period t and the old version is also available in that period, the price change recorded for the period t and period t + 1 indexes is determined, respectively, by the price relative of the old item for periods t and t – 1 and the price relative of the new item for periods t + 1 and t. The method does not require direct price or attribute compari- son of the old and the new products. Any difference in price in period t is attributed to item quality differences (Kokoski,1993:35). Because the overlap method is employed infrequently, its use (for within-outlet replacement item pricing) is unlikely to be a major source of quality bias in the CPI.10 The second option, explicit cost-based adjustment, is regularly used. Cost- based methods were applied to 11 percent of item replacements in 1995; they can be used when information about production cost differences between the replaced and the new items is available. Under the explicit cost-based method, the per-unit change in production cost, as reported by manufacturers, is subtracted from the 10It is fairly clear why the overlap method is used infrequently, given CPI price collection meth- ods. If the regularly priced item is available at time t and it is not known that it will be off the shelf in t + 1, there is no reason the commodity analyst would price a potential replacement at that time. Even assuming the eventual replacement was available in time t, without prior knowledge the method would require going back, at t + 1, to figure out the shelf price of a replacement at the time of the previous trip to the outlet.

EVOLVING MARKET BASKETS 119 change in the observed price paid by consumers. This method, frequently applied to new automobile models, rests on the strong assumption that the perceived value of improved features or new equipment is equal to the incremental costs incurred by the manufacturer to add them. If producers tend to exaggerate the cost of quality improvements, so that reported cost contains an element of pure price change, then the method imposes a downward bias on the CPI (Kokoski, 1993:35). In such a case, a cost-based adjustment erroneously attributes a portion of the observed price difference to quality change. Until competition catches up, however, one might expect profits to be made from an innovative product im- provement. In such a case, the price increase would be larger than the per-unit cost associated with the innovation. To the extent that the higher price equates with added value to consumers, the cost-based method (assuming accurate re- porting by the manufacturer) might, by ignoring the profit component, understate the quality component of the observed price change. The Boskin commission judged that BLS use of manufacturers’ cost data tends to underadjust for quality change and, in turn, imparts an upward bias to the CPI. The commission estimated quality change bias for new vehicles to be .59 percentage points per year for the period 1983-1996 (Gordon et al., 1997:86). The commission argued specifically that, in the case of automobiles, cost-based ad- justment did not include a number of manufacturers’ improvements that increased automobile durability and reduced production defects. However, Triplett argues that the Boskin commission assertions about quality bias for new cars were poorly informed (Triplett, 1997:27): Bureau of Labor Statistics (1997) listed changes, such as increased use of corro- sion-resistant metal, for which cost-based quality adjustments for automotive durability have been made in the CPI. Reduced defects must also have come about from changes made by the car makers. In my experience in the BLS, the auto manufacturers never overlooked quality changes when they submitted costs to the BLS. Rather, manufacturers tried to attribute too much price change to quality improvements. . . . The Commission’s idea that quality adjustments are systematically overlooked by the manufacturers when they make reports to the BLS is inconsistent with my experience with these data and also inconsistent with alternative evidence. Triplett adds that published hedonic studies of new automobiles have produced indexes that rise more rapidly than does the CPI cost-based adjusted index. Griliches (1971:11) also warned that “basing such adjustments largely on data furnished by manufacturers and on ‘producer costs’ may wind up overestimating ‘quality change,’ accepting as improvements expenditures which consumers may not interpret as such.” Deletion (and also class-mean, which is a more targeted variant of the dele- tion method) is used when the replacement and replaced items are judged non- comparable and when neither overlapped prices nor producer cost information is available. Excluding sample rotation, when new independent product samples are

120 AT WHAT PRICE? drawn, deletion is the most prevalent method used by BLS to address non-identi- cal item replacement in the index (Kokoski et al., 2000:3). If the outgoing item is last seen in period t –1 and the new item appears in period t, the old item will be used in index adjustments up to the one made from period t – 2 to period t – 1. The new item is used for the adjustment from period t to period t + 1 and thereafter. For period t – 1 to period t, change in the index component is proxied by the observed price change of other goods in the same CPI item stratum. The traditional (non-class-mean) deletion method assumes that the pure price change from the replaced to the replacement item is the same as that for the composite of all other goods in the class. Any remaining portion of the price change observed for the good in question is attributed to quality factors. The direction and extent of residual quality biases associated with this method are disputed, but they essentially hinge on whether the “true” quality-adjusted price change for the item that changed is greater than or less than the measured price changes of the same-class items that were used in the imputation (Triplett, 1997:29). Triplett and some BLS researchers have argued that deletion can inappropri- ately attribute a portion of price changes to quality change and, therefore, lead to overadjustment (downward) of price quotes. Their argument is based on the observation that manufacturers are, at least in some sectors, more likely to change prices when a new model is introduced.11 In an extreme case, if prices changed only when models did, a deletion-based index would pick up no price change at all. Triplett’s suspicion of a downward bias is corroborated by Moulton and Moses (1997), who demonstrate that a disproportionate number of price quote changes do in fact occur when new models or varieties of certain goods are introduced. It follows that, if the prices of unchanged models are the only ones that count, as is the case with the standard deletion, the method would impart a bias. Comparative hedonic studies have also indicated a downward bias associ- ated with deletion (e.g., Liegey, 1993). Prior to the Moulton and Moses (1997) work, the Boskin commission arrived at a different conclusion—that the bias is likely to be upward—stating that the deletion method “bases price change on models that are unchanged in quality and may be further along in the price cycle (Boskin et al., 1996).”12 It is worth noting that nothing precludes the coexistence of both the type of bias that Moulton/ Moses detected and the type that Boskin hypothesized. Research performed by the BLS indicates that producers frequently take the 11BLS research showing that price increases tend to coincide with the roll-out of new models is best documented for the apparel and upkeep strata; see Armknecht and Weyback (1989), Liegey (1994), and Reinsdorf et al. (1996). 12The underlying assumption here is that product prices drop, or rise less rapidly, immediately after a new product’s introduction into the market. This assumption has undoubtedly been true for computers and electronic devices in general but is less clear for other categories of goods and services.

EVOLVING MARKET BASKETS 121 opportunity afforded by introduction of a new model to piggyback price increases (U.S. General Accounting Office, 1999:77). The class-mean method was devel- oped and instituted by BLS to address the problem posed by this pattern. Like deletion, the class-mean method is used to impute the price of a changed item, but it does so from a set of similar goods further limited to those (1) classified as comparable replacements or (2) that could be explicitly quality adjusted by a hedonic or direct cost method. The underlying assumption is that price inflation is different for items that undergo replacement than it is for models that do not change. Looking at this restricted class of goods allows the price trend of a replacement item to be imputed from price quotes observed for other models that have turned over; but by limiting the set to items deemed comparable, it is hoped that quality-related elements are not a major factor. The method has been used in the new cars index since 1989 (Moulton and Moses, 1997:327).13 Assessment of Views on Within-Sample Quality Adjustment As noted above, the Boskin commission report did not seek to identify bias specifically associated with CPI procedures for handling quality change. This was largely a by-product of the commission’s decision to estimate new goods and quality change bias together, by CPI category, using independent evidence on quality-adjusted price changes. The work by Triplett (1997) and by Moulton and Moses (1997) indicates that full assessment of quality change should also include an examination of potential biases associated with BLS adjustment procedures. Summarizing the impact of quality adjustments applied by BLS item replacement methods, Moulton and Moses (1997:348-349) conclude: For certain important categories of items considered by the Advisory [Boskin] Commission, it would be difficult to argue that the CPI does not overstate the rate of price change. In other cases, however, any bias seems likely to be con- siderably smaller than the advisory board has estimated and, in certain cases, it could even be negative. . . . Our measurements of quality effects . . . show that any quality bias could go in either direction, either through inadequate quality adjustment (as emphasized by the advisory commission) or through excessive quality adjustment by the application of the link method to items with rising prices. In addition, Moulton and Moses demonstrate that quality adjustments by the BLS do have a significant effect on measured price change. Examination of BLS methods calls into question the Boskin commission’s view that price growth associated specifically with CPI sample items is biased upward. The debate over the extent to which the treatment of quality improvements produces upward bias in the CPI has been, to a substantial degree, a conflict 13We discuss the class-mean method in more detail in the section on hedonics (since class mean is usually what hedonic methods have replaced).

122 AT WHAT PRICE? supported by anecdotes. It is not the purpose of this panel to provide new esti- mates of bias or a detailed technical assessment of BLS procedures. However, the panel has identified several broad issues relating to quality adjustment that de- serve attention. First, how can the BLS best assure that the process whereby it identifies and measures quality changes is as objective as possible is not driven by highly subjective assessments of where the problems are likely to be and pays appropriate attention to areas where quality deterioration may have occurred? Second, while adjusting for quality change can in some cases be relatively straightforward, it usually involves product characteristics that are difficult to quantify. Airline deregulation, for example, led to generally lower airfares, but the low fares produced more crowded planes, more cancellations, and more fre- quent and longer delays (quality deterioration from the standpoint of travelers), as well as an increase in the frequency of flights between many pairs of cities (a quality improvement). How can these serious measurement problems be ad- dressed so that the value of these kinds of quality changes is reflected in the CPI? Many of the quality change examples used in recent critiques that find an upward bias in the CPI strongly suggest that quality improvements are over- looked. But many of the examples have been chosen from visible product classes presumed to bias the CPI upward. Furthermore, it is difficult to know where BLS should draw the line between adjustments that are sufficiently replicable to be used for producing a publishable index and adjustments that ought to be part of an ongoing research program but are not yet (and may never be) suitable for publi- cation by a federal statistical agency. HEDONIC REGRESSION METHODS Hedonics currently offers the most promising technique for explicitly adjust- ing observed prices to account for changing product quality.14 Hedonic regres- sions are used to estimate the value of specific bundles of individual characteris- tics that, when packaged together, form goods. The principle underlying hedonics is that, if consumers face observable relationships between goods’ characteristics and their prices, one should be able to use these relationships to disentangle pure price changes from quality changes. Hedonics essentially replaces the price of goods with the price of bundles of characteristics.15 14This sentiment dates back as far as the Stigler commission report (1961), and is reflected in recent work by Triplett (1990), Kokoski (1993), Boskin et al. (1996), Fixler et al. (1999), and many others. 15This basic idea is useful in a variety of other contexts. Particularly when considering product design, marketers routinely treat products as bundles of characteristics; see Green and Krieger (1985). And hedonic regression is routinely used in real estate appraisal and assessment: equations relating sales prices to the characteristics of properties sold during a particular period are widely used to predict the “missing” sales prices of properties that did not change hands; see, e.g., Kang and Reichert (1991).

EVOLVING MARKET BASKETS 123 It is essential to note that hedonic techniques expose a purely empirical relationship between prices and variation among different models of a good. The results of hedonic regressions can be used in either a COGI or a COLI frame- work. Zvi Griliches, who helped pioneer the application of hedonic methods to price index construction, commented in 1976—and cited the comment approv- ingly almost 15 year later (Griliches, 1990:189): What the hedonic approach attempted was to provide a tool for estimating “missing” prices, prices of particular bundles not observed in the original or later periods. It did not pretend to dispose of the question of whether the various observed differentials are demand or supply oriented, how the observed variety of models in the market is generated, and whether the resulting indexes have an unambiguous welfare interpretation. . . . Its goals were modest. It offered the tool of econometrics, with all its attendant problems, as a help to the solution of the first two issues, the detection of the relevant characteristics of a commodity and the estimation of their marginal market valuation. One potential advantage of hedonics is that a market may offer products that display a constant set of characteristics over time, even though specific models (and the corresponding characteristic bundles) change. Moreover, in some cases the link between what consumers ultimately value and product characteristics may be more intuitive than the link to a product itself. To this point, Griliches (1990:191) wrote: “Buried within the hedonic idea was the germ of Becker’s (1965) ‘household production function’ and the notion that one should look at the relevant activity as a whole, at its ‘ultimate’ product in terms of utility or produc- tivity, and not just at the individual components.” A hedonic function relates the price, pit, of variety or model i of some product in some period t, to a vector of its relevant characteristics, zit : pit = ht(zit).16 In the examples of butter and gasoline discussed at the start of this chapter, z consists of a single variable (ounces of butter and miles of driving, respectively), and there was an implicit presumption that h should be simply proportional to that variable. In more realistic cases, there are multiple relevant characteristics, and h is generally not a linear function of their values. In a typical hedonic regression, price, or the logarithm of price, is the dependent variable, and identifiable and quantifiable product characteristics serve as the explanatory vari- ables.17 In a well-specified equation, coefficients on the explanatory variables reveal the marginal relationship between the product characteristics and price at 16Econometric estimation of hedonic functions dates back at least to the work of Waugh (1928) and Court (1939). This approach received considerable impetus from the seminal work of Griliches (1961). 17Interaction terms and nonlinear transformations are also sometimes employed. Some models call for additional explanatory variables such as time period indicators, outlet type, or brand name that may not always be directly indicative of product quality. The implications of the latter additions are discussed below.

124 AT WHAT PRICE? various values of z. The basic idea behind hedonic techniques is that one can use a hedonic equation to calculate the expected price of a particular variety—which may not in fact be offered for sale in the period being considered—based on its observable characteristics. Then, as long as the set of observable characteristics includes all characteristics that matter to consumers and the equation is properly specified, these results can be used to correct for product quality change. To estimate hedonic equations, variation (either cross-sectional or longitudi- nal, depending on model specification) in the measurable quantity of an attribute is needed to produce coefficient estimates. Categories of goods for which quality change is frequent but incremental, and for which characteristic changes are easy to measure, are considered the best candidates for hedonic analysis. Most obvi- ously, products must have characteristics that are clearly identifiable as valued by consumers. For computers, a relatively easy case, these might include processing speed, hard drive space, memory, and monitor size. However, in most instances, quality characteristics are more difficult to identify, let alone quantify. For ex- ample, measuring the performance of cars is highly subjective, as is quantifica- tion of their handling, comfort, or safety. Apparel is even more difficult, since consumers’ valuations may change over time with fashions. Identifying the char- acteristics of services that consumers’ value can also be very difficult. The successful use of hedonic methods rests on a modeler’s ability to iden- tify and measure quality-determining characteristics and specify an equation that effectively links them to the prices of different models or varieties. It also de- pends on the availability of good data. In order to produce meaningful results, one generally needs data on more product models than are represented in a typical price index’s sample of items. In addition, the reliability of regression coeffi- cients depends directly on the amount of variation (both in terms of presence of indicator variables and magnitude of continuous variables) in the set of character- istics specified in the equation. In theory, quality adjustments to observed model prices (or of a product- specific index covering all models) can be estimated directly from the hedonic regression. In practice, the critical question is whether one can reliably estimate functions that capture the relationship between market price and characteristics that confront individual consumers. Here, the issue of consumer heterogeneity (see Chapter 8) arises again in a way that affects the index’s distributional prop- erties. First, without heterogeneity there would be no hedonic surface in the first place, since identical individuals would choose the same variety (bundle of char- acteristics) and pay the same price. But because individuals value product charac- teristics differently at the margin, quality adjustments can alter the relevance of an index as a representation of price changes faced by specific groups or indi- viduals. For instance, people who do not use cell phones do not care about their characteristics, and even the preferences among those who do use them vary greatly. Thus, when prices of cell phones are adjusted to compensate for quality change associated with model turnover, the overall index only becomes more

EVOLVING MARKET BASKETS 125 accurate (as a quality-adjusted measure of price change) for one segment of the population. It may become less accurate for those who do not use cell phones. Moreover, even if everyone faced the same cell phone prices and everyone used a cell phone, the fact that different people choose cell phones with different bundles of characteristics means (because the hedonic function h is not generally linear) that they generally face different marginal prices for characteristics. These issues become thorny when specific groups—such as the poor, the elderly, or those living in a certain area—consume “bias-corrected” products at significantly different rates or in significantly different varieties than do others.18 Note, though, that these problems are caused by consumer heterogeneity, not the use of hedonic techniques, though they surely complicate our understanding of what is being done with hedonic methods. The hedonic conceptual framework brings to light several difficult issues that have not been fully worked through either by BLS or by academic researchers. In the next section we compare competing hedonic approaches. Alternative Hedonic Methods Two basic hedonic adjustment techniques have been developed. For the most part, BLS has pursued an “indirect” approach, designed essentially to supplement price-linking methods when the quality (defined by observable characteristics) of outgoing and incoming models cannot be matched. The method uses a single estimated reference period hedonic function to adjust price quotes—for replace- ment items that appear at sample outlets—prior to their integration into normal index calculation. The indirect hedonic method is viewed by BLS as an alterna- tive to deletion or cost-based methods for adjusting prices and to the judgments of commodity analysts for assessing comparability of specifications (as such the merits of hedonic methods should be judged against these alternatives). Aca- demic economists, in contrast, have devoted more attention to “direct” tech- niques, in which indexes are produced directly from estimated hedonic functions based on data for both base and comparison periods. Indirect Methods BLS uses the term “indirect” in reference to a specific way of using hedonic functions to deal with situations in which one variety of a good tracked in the CPI 18As the economy moves toward greater product heterogeneity, this aggregation problem may potentially increase with or without hedonics. Noncomparable item replacement procedures gener- ally attribute price differences (or portions thereof) to quality difference, though many people would not be willing to pay that difference. Independent of the mix of quality adjustment techniques used by BLS, there is a positive correlation between the extent of changes in product characteristics and the magnitude of the aggregation problem. The magnitude of the problem, even without hedonics, is not necessarily an argument for ignoring changes in product characteristics.

126 AT WHAT PRICE? system—with a specific vector of characteristics z1 and price p1—disappears and is replaced by another—with “similar” characteristics z2 and price p2. This method is indirect because it involves adjusting, post hoc, the observed price difference between the outgoing and the replacement items based on the portion of the price change attributable to quality change. The magnitude of the adjustment is deter- mined by the estimated hedonic function and the differences between the charac- teristics bundles supplied by the old and new items. It is possible and, for reasons of data availability, often necessary to base these adjustments on a hedonic func- tion that is estimated with data from a period well before the substitution occurs. Though the difficult econometric problems that plague all hedonic analy- sis—e.g., identifying appropriate functional form and relevant product character- istics—complicate the indirect method, it has considerable commonsense appeal, at least relative to the alternatives: using the price relative (p2/p1) with no adjust- ment for quality change, assuming that the observed price change is due entirely to quality change, or adjusting for quality using one of the standard replacement methods. The fact that, among hedonics approaches, the indirect method is the least demanding in terms of data and procedure adds to its practicality. It simply requires using cross-sectional price and model characteristics data (similar to that which BLS already tracks) to estimate hedonic functions periodically.19 This function can then be used to estimate the price that would have been charged in the period studied for new models (with the same characteristics but different quantities of them) that are to be brought into the index (see “Technical Note 2” below). In contrasting alternative hedonics approaches, it is imperative to under- stand that the indirect method is applied by BLS in a comparatively narrow manner—to adjust price quotes, gathered under normal procedures, of replace- ments for items that have permanently disappeared from a sample outlet. For most products that are now hedonically adjusted (all by the indirect method), the monthly number of quotes adjusted is quite small, as is the effect on the monthly index for the relevant stratum. Direct Methods Two distinct direct hedonic adjustment approaches have been developed: the direct time dummy method and the direct characteristics method. In the direct time dummy method, data from multiple periods are used to estimate coefficients of a function relating the logarithm of price to a set of product characteristics and a set of 0-1 dummy variables for the periods covered.20 As discussed below, this 19Though the type of data required are similar, BLS typically has needed to expand its sample, or purchase commercial data, in order to generate a sufficient number of price points to estimate the hedonic models recently introduced into the CPI process. 20Work on the time dummy method has mainly been developed in the academic literature. Key studies include Griliches (1990), Triplett (1990), Berndt et al. (1995), and Arguea et al. (1994).

EVOLVING MARKET BASKETS 127 procedure assumes that for any two periods, t and u, ht(z) = Ktuhu(z) for all characteristics bundles z, where Ktu depends on t and u but not on z. That is, between any two periods the prices of all models (actual and potential) are as- sumed to differ by the same percentage. If this assumption is correct and the hedonic function is correctly specified, the characteristics variables pick up all price changes driven by quality changes in the menu of varieties on the market and coefficients on the time dummies pick up the residual pure price change. The index—interpreted as the price ratio net of the quality component captured by the characteristics variables—is produced directly from the difference in the time dummy coefficients from period to period. If the dummy variable for the base period is omitted, as is standard, the antilogarithm of the time dummy coefficient for any other period t gives the ratio of the price(s) of the good in question in period t to the price(s) in the base period.21 Similarly, the antilog of the difference between the time dummy coefficients for any two periods gives the price relative between those periods. Under the time dummy method, a single regression covering all periods must be run each time the index is produced. Since regression coefficients involving the characteristics are held constant across periods, changes in marginal cost ratios or in consumer demand patterns are assumed to be negligible. Thus, the basic relationship between product characteristics and relative prices (as well as the mix of characteristics available at market) must be stable in order to accu- rately isolate the price component associated with quality change over successive periods. This stability is what allows time dummy coefficients to be interpreted as the pure price effect.22 The key problem with the time dummy approach is that, for product areas in which quality change bias is likely to be an issue, the relationship between price and characteristics often changes rapidly. As an example, it is unlikely that con- sumers value, on the margin, a 10 percent increase in computer hard drive memory the same now as a year or two ago. If regression coefficients assumed to be 21Triplett (2001b:6-7) notes that the dummy variable method, when specified in a double-log or semilog functional form, produces a price index based on the geometric mean formula. Since statisti- cal agencies have begun moving toward using the geometric mean formula to construct elementary item indexes (for other reasons), time dummy approaches have become more consistent with the prevailing methodology. 22The problem of obsolete regression coefficients on characteristics is not unique to the time dummy approach. Certainly, the coefficients produced by the indirect approach, if not updated, are also susceptible to the same problem. However, the magnitude of the effect that the changing “true” relationship between characteristics and price can have on the index is more limited for the indirect approach. An index that is adjusted with the indirect hedonic approach will typically be less volatile because it is only affected by those variables representing characteristics whose values have changed from one period to the next. By contrast, all product characteristic variables that experience a diver- gences between their estimated and “true” relationship to price affect the time dummy coefficients and, in turn, any index derived from them.

128 AT WHAT PRICE? constant over time are in fact not constant, the estimated time dummies will reflect a mixture of pure price changes and quality changes, and the resulting index will be biased. More generally, there is neither theoretical support nor much empirical evidence for the assumption that prices of all varieties of particu- lar products generally move proportionately over time. The second direct approach, variants of which have been suggested by Pakes and Levinsohn (1993), Feenstra (1995), Diewert (2001), and others, is what we call the direct characteristics method. Under this approach one estimates a sepa- rate hedonic function for each period and computes price relatives for the product under study by, in effect, comparing the functions for the periods involved. The idea is not to require that all estimated nondummy coefficients differ between periods; it is rather to impose between-period coefficient equality only when that hypothesis withstands statistical scrutiny. To the extent that data from multiple periods can be pooled, estimation efficiency (always a concern in these studies) can be enhanced.23 In contrast to the time dummy approach, the direct character- istics index is—as its name denotes—constructed from the characteristics coeffi- cients, which are in general allowed to vary over time. The method also offers an advantage over the BLS’s deletion or indirect hedonics methods in that it allows for correction of any sample selection bias that may be created because price changes are only sampled from the set of goods or services that remain un- changed from period to period. However, despite its conceptual appeal, there are reasons that, given the current state of the art, the direct characteristics approach does not have broad applicability across CPI categories. One issue, which applies to all direct meth- ods, involves the general problem of price data that reflect nonobservable seller attributes. Outlet bias (discussed in detail in Chapter 5), for instance, is difficult to control for in an index produced from a time dummy regression or by relating hedonic functions for successive periods. In contrast to other methods in which prices for replacement items are quoted from the same outlet, product price and characteristics data are combined from multiple sources to estimate direct he- donic indexes (Triplett, 2001b:3). The most obvious obstacle to widespread use of direct hedonic methods, though, involves the data requirements and the operational difficulty of produc- ing characteristics-based indexes on a high-frequency, up-to-date schedule. To produce such an index, routine data collection and processing procedures would need to be directed toward monthly pricing of a comprehensive set of varieties, chosen to represent a population’s consumption, rather than a limited sample. Most importantly, it would be necessary to gather data on the sales of all impor- 23Two related direct approaches, both of which give the same result as the direct time dummy method when its assumption of stability of nondummy coefficients (and thus of proportional shifts in prices of all varieties) is satisfied are discussed below in “Technical Note 2.”

EVOLVING MARKET BASKETS 129 tant varieties with at most a one-period lag (see “Technical Note 2” below). Given current technology, estimating hedonic surfaces for, say, September 2000, in time for the release of the corresponding monthly CPI is infeasible for most goods and services. Due to the narrow range of products for which data requirements can be satisfied, even proponents of the direct characteristics method acknowledge its current limitations. Pakes recommends starting with computers and moving slowly into other areas. Writing about the applicability of hedonics to index construction, he cautions (Pakes, 1997:9): There are, of course, several detailed decisions which will have to be made before the statistical agencies could produce hedonic adjustments for a set of subindices (among them a decision on the instances in which the hedonic bound is likely to be suspect). Moreover, any shift to hedonics will require prior exper- imentation by commodity group, and will generate adjustment costs. These warnings aside, moving in this direction would not require a complete overhaul of BLS pricing methods. In fact, data requirements for the direct charac- teristics method fit in fairly well with current collection procedures since charac- teristics must now be tracked in order to judge comparability for replacement situations. Even to improve the CPI under current methodology, better quantity and characteristics data are needed, which is what would also be needed here. In addition, the current requirement that identical products be found at outlets by BLS field agents over index periods might be relaxed since only characteristics— which might be found on a number of similar products—need to be tracked. Silver (1999:19) suggests that agency data collection needs for hedonic in- dexes might be met in the future with panels of consumer data for frequently purchased items and scanner data for durables. Paasche, Laspeyres, and superla- tive formulations could be produced, assuming that the time needed to process comprehensive product scanner data is short enough to allow for base and current period weights to be constructed. BLS Application of Hedonic Methods The Boskin commission attributed more than half of its estimated 1.1 per- centage point CPI bias to a failure of the index to fully account for changing product quality and the appearance of new goods. The BLS has responded to recommendations by the Boskin commission and others (both before and after Boskin) to address this perceived flaw by expanding its use of hedonic quality adjustment—specifically, the indirect method—in the CPI. Kokoski et al. (2000:3) characterize the hedonic method, or class of methods, as the “currently preferred method of quality adjustment.” The BLS position is that hedonic analy- sis provides meaningful information for inferring the value consumers place on quality change and that hedonic function estimates based on regression analysis can be reliably used for certain items to make quality adjustments to indexes.

130 AT WHAT PRICE? Hedonic methods were first applied in the CPI during the early 1990s to apparel categories. Initially, the technique was used to develop criteria for identi- fying appropriate comparable replacements for disappearing items. Shortly there- after, it evolved to its current use as a tool to filter out the quality-driven compo- nent of observed price changes associated with item replacements. For apparel and most other items, hedonics is typically used to make one-time quality adjust- ments (using the indirect method) when new items replace outgoing ones.24 Since January 1999, hedonic analysis of computers and televisions has been incorporated into the official index. The TV index was considered a good candi- date since models undergo frequent but nonradical quality change and also be- cause TVs occupy an entire item stratum. The latter feature is convenient because it allows the elementary index to be constructed without combining separate substratum item indexes calculated using different methods. For TVs, a semilog functional form is used to adjust observed prices to account for variable quality characteristics, such as screen size, wide screen, display, projection, and surround sound. Equations must be respecified as new features appear, which become new explanatory variables (Fixler et al., 1999:10). An indirect approach, as described above, is used. BLS has also tapped into research done for the producer price index (PPI) to develop hedonic regressions for large-scale and network as well as desktop com- puters. Hedonic adjustments have been used in the PPI since 1990, and applica- tion to the desktop computers stratum of the CPI was incorporated in January 1999. Explanatory variables include such characteristics as chip speed, system memory, and monitor type and size (Fixler et al., 1999:11). Since computer technology changes rapidly and the relationship between computer features and value appears unstable, regressions have been respecified frequently (every 3 to 12 months). For 1998, the first year that hedonic regressions were used to adjust computer prices, the price index for personal computer and peripheral equipment was reduced by 6.5 percent, relative to what it would have been had the new method not been implemented. A hedonics technique has also been applied to the “rent of primary resi- dence” and “owners’ equivalence” components of the index since 1988. A more restrictive type of the indirect hedonic method is used to estimate only the effect of aging on the value of housing units. In the application, rents are adjusted for age using a nonlinear (age and age-squared variables) specification. Items Targeted by the Recent CPI Hedonics Initiative The BLS is currently conducting research that will extend the use of hedonic regression models to additional CPI items. Kokoski (1993:12) states that “for 24Interestingly, over the last six months of 1991, using hedonics raised the rate of price change for the apparel category by 0.4 percent per year (Liegey, 1994).

EVOLVING MARKET BASKETS 131 many CPI components, a hedonic approach will likely be adopted before the next scheduled revision in 2002.” In fact, the move to expand the hedonics program has already accelerated in response to the fiscal year 1999 CPI improvement initiative.25 The initiative has added hedonic price adjustment to the following item groups: • microwave ovens (effective July 2000) • refrigerators (effective July 2000) • camcorders (effective January 2000) • VCRs (effective April 2000) • DVDs (effective April 2000) • audio products (12 products, effective January 2000) • college textbooks (effective July 2000) • washers and dryers (effective October 2000) Specific CPI strata are chosen for the expanded hedonics program using the following criteria (Fixler et al., 1999:13): • There is a perception that the standard procedures inadequately account for quality change (it is unclear if BLS is focusing on items for which quality is perceived to be changing or specifically increasing). • There is a belief that hedonic models could be developed for at least a subset of items in a stratum (presumably beliefs must be supported by both data availability and theoretical considerations). • A significant percentage of price quotes exist for substitute items relative to the total number of price quotes for the stratum. Fixler also notes that the list of candidates provides nice contrast in terms of placement within item life cycles. At one end, DVDs are new and have recently undergone rapid technological development while, on the other, refrigerators and microwaves are well into the product cycle and technology is comparatively stable. It is not surprising that the goods included in the initiative are from the “appliances including consumer electronics” category since it is a product area that has undergone highly visible change. (This category accounted for the larg- est share of the Boskin commission’s estimated 0.6 percent unmeasured quality and new goods bias.) Much of the ongoing and proposed hedonics-related research must be sup- ported by additional data collection, since routine CPI sampling procedures often 25The hedonics project was one component of the 1999 CPI improvement initiative, which also designated funds to expand the sample size of the Consumer Expenditure Survey (CEX), to quicken introduction of new products and CPI market updating, and to develop new superlative indexes (Liegey and Shepler, 1999:34).

132 AT WHAT PRICE? yield an insufficient number of models to permit reliable estimation of realisti- cally complex hedonic functions. The initiative specifically provided funding to collect additional price observations from current CPI outlets (2,500 quotes dis- tributed among eight items). For some experiments, BLS field agents are also collecting observations from added outlets (as is the case for camcorders); for others (audio products), market data have been purchased from vendors such as A.C. Nielsen or NPD.26 It is important that BLS continue to examine the implica- tions of using non-uniform data for estimating hedonic regressions and for index construction generally. The CPI Hedonics Model Most of the recent BLS work uses the indirect adjustment method. Price adjustments are calculated with an equation in which the (logarithm of) price is estimated as a function of product characteristics. The portion of the observed price difference between a replaced item and a substitute item assigned to differ- ences in quality is determined by the differences in characteristics variables and the associated coefficient values. The process of specifying the model typically involves researching consumer magazines and manufacturer and industry infor- mation to develop a sense about which characteristics are related to price. Several specifications may be experimented with prior to final model selection. In the case of VCRs, BLS’s final specification consisted of all dummy variables on the right-hand side, each indicating the presence of a particular feature (number of video heads, auto rewind, hi-fi stereo, etc.). Liegey and Shepler (1999:27) write: “The specification for the final VCR regression model was deemed satisfactory, primarily because the magnitude and direction of the parameter estimates matched a priori expectations. The high R-squared value further validates the model.” For several of the applications, the dependent variable is the list price, not the trans- action price. When the model is estimated using retail list price as the dependent variable, a dummy variable indicating that the item was sold at a sale price is included in the model to capture the (negative) effect on actual price in the data.27 In addition to tangible characteristics, brand dummies are often included as explanatory variables. Inclusion of brand names in the equations has been de- 26The audio project, which relies on purchased point-of-sale data, and the video project, which relies on conventional in-house surveys, may provide useful contrasts. The audio data include price and units sold but have limited information about attributes, which forces BLS to supplement the data with manufacturer specifications. Typically, collecting vendor data is more expensive than collecting a sample internally, but the data are available with greater frequency. 27This is often done because of data constraints. Earlier studies, such as Liegey and Shepler (1999) on VCRs, used Consumer Reports and not CPI price data. Mary Kokoski has questioned this practice commenting that, “since no one really pays full prices, do they (the results) really reflect the equilib- rium assumptions that underlie the hedonic method?” (Liegey and Shepler, 1999:32). The panel shares Kokoski’s concerns.

EVOLVING MARKET BASKETS 133 fended on the grounds that coefficients for the variables were robust, that it increased the explanatory power of the models, and that is does not create multicollinearity problems.28 Nonetheless, including brand names is controver- sial. It is reasonable to worry that the brand variables may “steal” effects from other characteristics—both those that are and that could be included in the model—and thereby bias the estimated effects of characteristics on price. If one assumes that brand, in itself, does not lead to higher valuation by consumers, one must believe that it is an acceptable proxy for unmeasured quality characteristics. Incidence of repairs might be one such example. However, the BLS study on microwave ovens found that the brand names most valued by consumers were actually those with the highest incidence of repair (Liegey, 2000:5). Moreover, brands are repositioned in terms of relative quality from time to time, and reputa- tions sometimes change in response to advertising campaigns, so that brand dummy coefficients may be inherently unstable. Given the difficulty of interpret- ing coefficients on brand variables, it would be instructive if researchers docu- mented their results with and without brand variables and provided a hypothesis as to what aspects of product value the brand variables are capturing. The BLS hedonics research program has helped reveal that, in practice, applying hedonic methods to price indexes involves confronting very tough is- sues. Characteristics cannot be chosen in a formulaic manner—lots of ad hoc judgments are inevitable—and, once chosen, estimated coefficients may exhibit implausible signs.29 Furthermore, models need to be regularly updated because the relationships between characteristics and price are not stable for long periods. For instance, Liegey and Shepler (1999:27) show that new features on VCRs have a large predictive effect on price but, as they become common, their impact quickly recedes. This is a good example of the kind of work BLS must continue to undertake to support expansion of its hedonic program. Investigations into model stability for different product areas are much needed to improve judgments about the frequency with which hedonic regressions should be reestimated. While the panel believes that the BLS research program is essential to im- proving understanding of the theoretical uncertainties about hedonic methods, our concerns have not been allayed by what has actually been done so far. Given these ongoing concerns, we are still quite uncomfortable with extending the 28Moulton et al. (1998) include indicator variables for brands in their study of televisions. They argue that brand name is important since “a set with the same screen size and other observable characteristics with a premium brand name, such as Sony, may sell for as much as 50 percent more than a similar television from a less prestigious brand” (p. 9). The authors acknowledge that, if additional characteristics could be added to the regression equation, the effect of brand variables might be reduced. 29Pakes (2001) has argued that, given rapid entry and exit and great product differentiation in technologically innovative markets, it may not always be clear what the “right” sign on a characteris- tics variable should be.

134 AT WHAT PRICE? application of hedonic models, in their current state of development, to additional index categories for use in the CPI. Yet the panel is not convinced that anyone could have done this work any better—or is better equipped to continue work in this area—than BLS. Hedonics Use and the CPI To make clear the implications that the shift from implicit quality adjustment to hedonics may have on the CPI, we must first describe the BLS in-store item replacement procedure.30 When a routinely priced item becomes permanently unavailable, BLS field agents are instructed to choose the most similar alternative available at the outlet on the basis of a hierarchical list of characteristics specifi- cations. As explained above, the pure price change for a replacement item in the month of its introduction in the index is measured as the average price change that month among similar items (selected according to one of several different methods). Any remaining difference between the price of the replacement item and the old one is assumed to reflect quality change. The practice of choosing the “most similar” item as the replacement means that the potential quality difference between outgoing and incoming items is smaller than if the practice were to select the most advanced, the newest, or the most frequently purchased product within the same class. Also, this practice increases the number of substitutions that are deemed comparable and that do not require quality adjustment and reduces the magnitude of noncomparability for items that are price adjusted. In 1997, 58 percent of the almost 29,000 nonrent substitutions were judged comparable by commodity analysts (U.S. General Accounting Office, 1999:13). For the subset of substitute items that are deemed noncomparable, BLS then attempts a direct quality adjustment, using hedonics or cost-based calculations or a traditional indirect adjustment method. Other than in its application to personal computers, hedonic adjustments are producing little if any effect on the CPI.31 The effect of increased use of hedonics is limited by: • its narrow application to noncomparable substitute price quotes, • the nature of CPI item substitution itself, and 30U.S. General Accounting Office (1999) provides a detailed explanation of the rules that guide item replacement by BLS commodity analysts. The report also describes how the class-mean and other link methods of adjustment work. 31It should be noted that, even with hedonic adjustment, the rate of price decline for personal computers in the CPI or PPI is generally much smaller than that estimated by outside researchers. The research in this area is quite controversial.

EVOLVING MARKET BASKETS 135 • the fact that hedonics methods are being applied only to items that were previously quality adjusted by other BLS techniques. CPI hedonics models are only used to adjust a subset of substitute price quotes: that is, to control for quality change when a product that has disappeared from the shelf must be replaced by a noncomparable one. Hedonic adjustments have not been used to offset quality differences associated with item turnover generated by outlet rotation or with respecification of the CPI basket. This is an important distinction because implicit forms of quality adjustment (such as dele- tion) are already a feature of the in-store item substitution process. As currently used, hedonics is simply replacing another method of quality adjustment, and the item-by-item effect on index growth has been minimal and its direction ambigu- ous. In some cases, BLS’s hedonic models implied price adjustments that would have been larger than the standard (deletion) adjustment actually used; in others the adjustment would have been smaller.32 A more broadly based application of hedonic techniques—one that extended beyond routine item replacement cases caused by sample attrition to one that, for example, was also applied to price changes associated with new models appear- ing as a result of sample rotation—would be expected to have a larger effect on the index. For example, Moulton et al. (1998) simulated an index for televisions that did include hedonic comparisons of items that entered the CPI through sample rotation. Their analysis, which tracked product characteristics over a 5- year period, resulted in a much larger downward adjustment than a simulated index that applied hedonics only to in-store item replacements. The authors argue that sample rotation may be particularly important for TVs since, unlike comput- ers, when models with new features appear on the market, older sets of character- istics (models) remain available for a long time. This means comparable replace- ments can continue to be found, and commodity analysts need not turn to more radically changed models. Cutting-edge models, even those that quickly gain in market share, seem more likely to enter the CPI when outlet and item samples are rotated (Moulton et al., 1998:12). Moulton et al. (1998) recommend developing hedonic adjustment and data collection techniques that would make it possible to apply hedonics methods when new products enter the sample during outlet rotation. They also suggest changing the item replacement rule to have field agents select items that are more 32This evidence does not speak to questions about the extent to which adjustments are becoming more accurate; matched model (replacement or class-mean) adjustments may conceivably overstate or understate the impact of quality change on price and, while hedonic regressions produce addi- tional evidence about this, the value of the additional information is dependent on the validity of the model and the quality of underlying data used in its estimation.

136 AT WHAT PRICE? representative of prevailing quality choices in current consumption, then adjust- ing the selected substitutes with the standard indirect hedonic models.33 Since hedonics currently only replaces other procedures, its adoption by BLS has not led to more widespread adjustment for quality change in the CPI. Hedonic applications have most frequently taken the place of the class-mean method. As discussed above, the class-mean approach, like the deletion method, infers qual- ity differences by comparing the observed price change of the replaced and replacement items to the price change of other goods. However, compared with deletion, the pure price change is imputed from a smaller set of quotes. Quotes are still drawn from a specific entry-level item (ELI), index area cell, but only those for items that are comparable replacements or directly quality adjusted are included to calculate average price change for the class. The idea is that the price change of items that are comparable to what they have replaced will reflect only pure price change. This approach is designed to recognize the fact that sellers often use the occasion of introducing model changes to raise prices. The esti- mated monetized value of quality change is the residual of the observed price change of the substitute item after the average price change of similar but compa- rable replacement items has been accounted for. The class-mean method was instituted during the late 1980s partly to capture price increases that accompany introduction of new models. It is the designated method for item strata that experience frequent model and product line turnover. The deletion method misses these price increases because it only follows price movement associated with unchanged models. For this reason, Triplett (1997), Shapiro and Wilcox (1996), and BLS researchers have suggested that the method can overstate quality change and bias the index downward. The class-mean method, by imputing price change only from comparable replacements, attempts to account for the possibility that inflation is different for models that are chang- ing in comparison with models that are stable. But if producers are systematically more likely to include “pure” price increases (or decreases) for new models that are substantively different from old ones than they are for those with only a change in model number (as is the case with 60 percent of comparable VCR substitutions), then this correction may still understate pure price change. In this case, the class-mean method may still lead to a larger quality adjustment than a 33The issue of whether or not the current item replacement rule—choosing the closest comparable item—makes sense is a separate but important one. In quickly evolving technology areas, if the replaced item became obsolete, it is likely that the closest substitute is also near obsolescence. If, instead, BLS agents selected the newest model or the one with the highest sales, the frequency with which item substitutions must be made could possibly be reduced. However, such a change in procedure would require making larger quality adjustments, which might pose other problems, par- ticularly if one is not confident that currently available methods can really disentangle pure price and quality contributions to the observed price. A problem with using the newest model is that it will sometimes pick up features that do not last—not all innovations survive in the market.

EVOLVING MARKET BASKETS 137 perfectly specified hedonic adjustment. But hedonics equations cannot be per- fectly specified, and BLS research indicates that, in practice, the adjustments fall on both sides of the class-mean imputation. We look at some of these research results next. The Recent BLS Hedonics Initiative BLS’s hedonics studies are designed to produce equations that can be used to adjust noncomparable replacement item price quotes. Several of the studies com- pare the performance of a hedonically adjusted index against one simulated with the same methods as the published index. In as many cases as not, the hedonically adjusted price index increases at a faster rate than does the published (class-mean based) version. In other words, the hedonic quality change price adjustment is often smaller than the conventionally used implicit quality adjustment. For in- stance, for the period June 1999 through December 1999, substitute VCR price quotes used in the published index decreased on average by 4.3 percent (or, annualized, by 7.4 percent). For the same 7-month period, the hedonically ad- justed price for substitute quotes decreased by an annualized 2.2 percent (Thomp- son, 2000:6). This translates into 13.2 and 11.5 percent annualized decreases, respectively, for the published versus hedonically adjusted indexes for the other video equipment strata, of which VCRs are one subcategory. The study of VCRs by Liegey et al. (1999) showed similar results using 1997 data: hedonically adjusted price quote substitutions also grew less rapidly than the published in- dexes for refrigerators and audio equipment. In contrast, some of the experimental hedonics applications have slowed index growth by more than the implicit adjustment methods. The TV index falls more rapidly with hedonic adjustment. Moulton et al. (1998) produced a hedoni- cally adjusted index (of the type that has been adopted for other items in the CPI) for the period 1993-1997 that grew 1.4 percent less than the actual CPI for televisions that used the linking methods (p. 11). Table 4.2 summarizes the ef- fects on the CPI from a selection of recent hedonic applications. For many CPI items, the number of substitute quotes that are available to quality adjust is not large enough for the hedonic adjustment to seriously affect the strata index, much less the overall index. Moulton et al. (1998) note that confining hedonic adjustments to cases of noncomparable substitutions for any- thing other than very high turnover products like computers will not produce many significant effects on the CPI component indexes. Another factor that may affect index growth is the frequency with which items are deemed noncomparable and, hence, eligible for hedonic or class-mean adjustment. In each of the BLS studies, the breakdown of substitute items into comparable and noncomparable categories changed with adjustment mode. For instance, in the microwave oven study, moving to hedonics resulted in an in- crease in the number of noncomparable substitutes increasing from 5 (of 39

138 AT WHAT PRICE? TABLE 4-2 Major Effects on CPI Indexes from Five BLS Hedonic Studies Change in Average Price Change Number of Stratum Index for Product Substitutions Monthly Substitutions Hedonically Hedonically in Product Product Stratum Published Adjusted Published Adjusted Group VCR Video –13.2 –11.5 –7.4 –2.2 19 excluding TVs Audio Audio products productsa –7.4 –6.0 n.a. n.a. n.a. Refrigerators Major appliances 1.2 1.2 2.8 3.9 5 Microwaves Major appliances 1.1 0.8 5.4 1.7 4 DVDsb Video excluding n.a. n.a. n.a. n.a. 1 TV NOTE: Index and price changes shown in percent and at annual rates. aNot a CPI stratum. bToo few price quotes to estimate an adjustment to the overall stratum index. quotes) to 21. For VCRs the noncomparables decreased from 63 (out of 130) to 47. BLS explains why using a hedonics system might identify different non- comparable cases from a given set of substitute quotes (Thompson, 2000:5): For the purpose of calculating the quality adjusted index, all the substitutions were reevaluated. One of the benefits of using a hedonic model in evaluating substitutions is that the analyst has an opportunity to review price data and item characteristics with a statistical tool, thus enabling him/her to render judgments based on statistics rather than expert judgment alone. When BLS uses hedonics, comparability of substitute quotes is judged not in terms of an a priori determination about the amount of quality change, but by the extent to which price change is predicted by the regression equation.34 For substi- tute price quotes, “differences in the specification or characteristics data of the old and new items were identified to see if the parameter estimates in the hedonic model could be utilized to quality adjust the official price change” (Thompson, 34The “standard” comparability decision in forced when a commodity analyst must add to the sample a replacement item that does not match the detailed description of the old one. The analyst judges comparability on the basis of an examination of any differences revealed by the checklists for the old and new versions of an item.

EVOLVING MARKET BASKETS 139 2000:10). Of course, the way a model is specified affects where the line of comparability will be drawn, and it has not been established that the statistical procedure is an improvement over the standard process, particularly if the he- donic regressions are estimated by analysts who know little about the products being adjusted. It is essential that the expert judgment of commodity specialists be retained and perhaps supplemented with that of other product and marketing specialists. The fact that hedonic methods have produced item comparability judgments that are different from those made in conjunction with deletion methods is a concern (relevant to either method) for several reasons. First and most obviously, different judgments about comparability lead to different rates of quality adjust- ment, which has obvious implications for index performance—more price adjust- ments translate into a decreased rate of index growth. Second, the empirical work produced from the recent initiative indicates that price change associated with the comparable substitute quotes can also be significantly affected by the choice of quality adjustment technique. For instance, the 67 directly compared (non-ad- justed) VCR price substitutions produced for the published class-mean adjusted index decreased by 3.72 percent; the 83 directly compared price substitutions left over after hedonic adjustments were added decreased by only 1 percent. It is not intuitively obvious why the choice of noncomparable quality adjustment method should have such a large effect on the price change of the subset of comparable substitute quotes. Documentation supporting BLS research does not adequately explain the effect on comparables that results from switching quality adjustment methods. These criticisms aside, the move toward supplementing judgment—both for- mal and informal—with replicable, systematic methods of comparing non-identi- cal items is a move in the right direction. Econometric analyses of data that indicate how characteristics are correlated with price change have the potential to improve the ability of BLS analysts to determine what is and is not a comparable substitution. Even if a particular hedonic study is not convincing enough to be used for quality adjustment, it may still offer insights that improve analysts’ informal decision making. Statistical audits provide evidence about the variance that arises when different researchers, using the same data, attempt to replicate quality adjustments (Triplett, 2001b:9). Summary The incorporation of recent BLS hedonics research into the CPI has not produced evidence for the conclusions offered by the Boskin commission and others about the extent to which quality change biases the CPI when used as an approximation to a COLI. The research indicates, at least, that the commission underestimated the effect on the index of implicit quality adjustment measures already in place. Moreover, even a substantial expansion of hedonics, used as the

140 AT WHAT PRICE? BLS now does in the item replacement process, would not be likely to have a big effect on the CPI. Hedonic adjustments tend to wash out relative to those pro- duced by the implicit adjustments that they replace (the computer index is the exception). Confining hedonic adjustment to cases of noncomparable substitu- tions for anything other than very high turnover products is unlikely to signifi- cantly affect CPI component indexes. Also, current BLS rules for replacing dis- appearing products further minimize quality differences between outgoing and incoming products which, in turn, lessens the importance of which type of quality adjustment is ultimately selected. However, as the Moulton et al. (1998) TV study suggests, the application of hedonic adjustments in a different way and on a larger scale might produce more significant downward adjustments. The panel believes that the BLS should proceed cautiously in its efforts to integrate hedonics into the CPI. Further research, testing, and evaluation of hedonic methodology and specific applications should precede expansion of its use, such as to sample rotation—something that the panel is not in principle opposed to—where the impact on index growth would likely be more significant. CAUTIONS AND RECOMMENDATIONS Hedonic methods are not a cure-all for indexing problems related to quality change. Regression techniques do not deal with increases in product variety (e.g., of fruits and vegetables during the winter); nor do they help much with the problem of truly new goods (e.g., cellular phones). The main thing to be said for hedonic methods is that there is nothing better for dealing with certain aspects of the quality change problem. This is not an elegant defense, but it is a powerful one. To a large extent, this reality shapes our recommendations in this area. BLS should systematically investigate quality change across CPI compo- nents. Recommendation 4-1: In addition to its targeted intuitive approach (in which BLS selects for adjustment items thought a priori to have undergone quality change), BLS should pursue experiments to ana- lyze quality change in randomly selected items in order to increase the probability that within-sample quality change biases—both up- ward and downward—will be identified. Currently, hedonic regres- sion analysis is the leading candidate to serve as the main analytical tool in such experiments. One issue that will have to be addressed in such a program is the level of detail that is used in the item selection process. Selection could be randomized across the broad 211 item strata, at more detailed ELI levels, or somewhere in between. Whatever level of disaggregation is chosen, it is logical that selection probability should be proportional to expenditure (perhaps adjusted to account

EVOLVING MARKET BASKETS 141 for the rapidity of item replacement) and not random over items independent of weight. The issue of how to assess service expenditure categories will also pose special problems. One can imagine that the quality of various consumer services changes substantially over time, and certainly not always for the better—think of airline travel for instance. In principle, methods such as hedonics that are used to identify and adjust prices of quality-changed goods can also be applied to ser- vices; in practice, for many services, the problem of how to define output appro- priately must first be solved. The above recommendation identifies one element in what should be a broad-based hedonics research program. Recommendation 4-2: BLS should continue to expand its experi- mental development and testing of hedonic methods and its support of relevant outside research. This research should not be confined to that relating to price adjustment but should also examine the role of hedonics in statistical audits of the other BLS quality adjustment methods and in the review of replacement item selection procedures and comparability decisions. The above recommendations do not suggest that BLS should immediately expand the use of hedonics in constructing component indexes for its flagship CPI. In fact, the panel takes the opposite position. Recommendation 4-3: Relative to our view on BLS research, we recommend a more cautious integration of hedonically adjusted price change estimates into the CPI. This recommendation is based on theoretical considerations, not on empirical grounds. As documented above, the recent BLS expansion of hedonic price ad- justments to appliances and electronics has not had a large impact on those item subindexes. The current hedonics program, which only replaces other quality adjustment techniques, actually has an ambiguous effect on index growth. Thus, for practical purposes, the apparent rapid expansion of the use of hedonics is not a pressing empirical concern for those interested only in the accuracy of the final CPI numbers. Our conservative view on integrating hedonics techniques has more to do with concern for the perceived credibility of the current models. While there is an established academic literature on estimating hedonic functions, researchers are much less experienced using them across a wide variety of goods in price index construction. Thus, while members of the panel agree that BLS and others should continue to research the viability of hedonics, the methods may, in their current state of development, only be justifiably applied to a narrow class of goods. The list of unresolved econometric specification and data issues that may inhibit fully informed use of hedonic quality adjustment is a long one.

142 AT WHAT PRICE? For many classes of goods—and perhaps especially services—it can be ex- tremely difficult to identify which characteristics are actually associated with price. Despite the early success of hedonics to move quality adjustment in the CPI toward a statistically based approach, considerable judgment by researchers is still required. For instance, early introduction of video memory as an explana- tory characteristic in regressions for PCs yielded “implausibly high coefficient values,” so the variable was left out of initial specifications. Later the values “settled down and behaved much more reasonably” and the feature was included in the specification (Fixler et al., 1999:11). Likewise, in the hedonic model for TVs developed by Moulton et al. (1998:10) the “stereo sound” indicator was dropped because it predicted a negative (though insignificant) impact on price. Given the short history of this type of research at BLS, it is not clear what the benchmark should be for assessing what is or is not reasonable. Strange-looking variable coefficients could be indicative of larger problems—including omission of key value indicators, characteristic mismeasurement, and functional form issues. Whether for a standard comparability decision or for hedonic modeling, identifying and quantifying relevant characteristics is tricky when quality is tied to consumer perceptions that may not be constant over time. Also, it is next to impossible to collect full and timely information on certain types of product characteristics. Quality of fabric in clothing, for example, is determined by a complicated combination of characteristics—not simply by material type, but also by threads per inch, type of weaving, quality of dye, etc. Given the changing nature of fashion, a characteristic may be viewed as a negative quality at one time and as a positive quality at another. For instance, the original move from cotton to synthetic shirts was considered a quality improvement—but so too was the move back to cotton. Once identified, it is not necessarily any simpler to measure the characteris- tics thought likely to affect price: consider stylishness in clothing or handling in cars. Even for the best candidates, such as computers, attribute measurement can be problematic. For instance, how does one quantify the user friendliness of hardware or software? For most products, certain elements that contribute to its value will always be difficult to measure consistently. Theory provides little guidance to help determine the appropriate functional form for hedonic equations. Experience suggests that characteristics often inter- act in complex ways to affect value. When characteristics work in combination, nonlinear functional forms, perhaps involving interaction variables, must be used to produce reasonably robust results.35 Furthermore, when one product works as 35Curry et al. (2001) summarize some of the advantages of flexible functional forms (and even neural networks) in the context of hedonic modeling applied to consumer goods. The authors use detailed scanner data to estimate and test hedonic models with interaction effects for the U.K. televi- sion market.

EVOLVING MARKET BASKETS 143 a complement with another (e.g., hardware and software), it is conceptually un- clear how to quality adjust each in isolation. Nor does existing theory say much about the importance, or lack thereof, of explanatory power. It is hard to know when a hedonic function is good enough for CPI work: the absence of coeffi- cients with the “wrong” sign may be necessary, but it is surely not sufficient. When product and process innovations occur, tastes change, or input prices shift, hedonic surfaces may change rapidly. The ability of the BLS or any other agency to capture those changes in real time is, at best, doubtful. It is unclear whether usable estimates of hedonic surfaces can be routinely and rapidly com- puted for a wide variety of goods. For many goods, the relationship between characteristics and observed price may not be stable, and the best-fitting func- tional approximation may change across products or time, particularly when technological change is rapid. Research into the stability of coefficients for dif- ferent product groups is essential for making informed decisions about how often to reestimate hedonic functions. Without this information, reestimation schedules may be dictated by budget or other factors, which might result in outdated adjust- ments and be worse than doing nothing. If the hedonic functions were known in every period, some variant of the direct characteristics method would be the best way to derive price ratios. Some- times this would reduce to the direct time dummy method, but there is no reason to think this would occur frequently. Since the time dummy method has similar data requirements as the direct characteristics method but rests on much stronger assumptions that lack theoretical support—most notably, stable marginal impact of characteristics on price over time—the former has little to recommend it in principle. The time dummy method seems particularly unsuited to index use in rapidly changing product areas for which, presumably, quality adjustment is most warranted. Recommendation 4-4: BLS should not allocate resources to the di- rect time dummy method (unless work on other hedonic methods generates empirical evidence that characteristic parameter stability exists for some products). The biggest obstacle inhibiting use of the direct characteristics approach is that the data and analysis requirements are daunting. However, the payoff from using this approach could be substantial. The methodology can, in principle, produce quality-adjusted indexes that take into account changing marginal rela- tionships between characteristics, weighted by expenditure shares, and price. And, relative to the indirect method that adjusts an observed price change on the basis of individual coefficients, directly produced hedonic indexes are based on the entire hedonic surface which, in theory, generates more robust and precise estimates over different specifications.

144 AT WHAT PRICE? Recommendation 4-5: BLS should experiment with the direct char- acteristics method, beginning with a few, carefully selected goods. The timely availability of relevant data should be a key selection criterion. Though its statistical properties require more in-depth study, the indirect method seems at this time the most broadly applicable hedonic approach for use in the CPI. Recommendation 4-6: BLS should continue to study the value of the indirect method for a wide range of goods. A large part of its promise rests on the comparatively modest maintenance and data requirements relative to the direct methods. Different adjustment methods imply different updating intensities. Under direct methods, the hedonic regres- sion is part of index construction and must be rerun each time the index is recalculated. Thus, data on prices, characteristics, and purchase shares of a large set of varieties are required in each period and such data must be in hand for the current period before the hedonic function can be estimated and the index com- puted. Under the indirect approach, results can be obtained with only periodic re- estimation. Only past period estimates are required, so there is less time pressure on data collection and analysis. However, when the relationship between charac- teristics and price moves quickly, even the indirectly used hedonic functions must be reestimated and, when characteristic sets change, they have to be re- specified if they are to remain accurate. Ideally, regression equations would be updated every month. Practical considerations all but eliminate this possibility; BLS is not equipped or adequately funded to do this on a large scale. Rerunning current models with new data may not be overly burdensome, but respecifying models is highly labor intensive. Given that data collection and model estimation requirements may impose more than a 1-month lag in many cases, it may be necessary to figure out how best to use an estimated surface based on 6-month- old data to compute hedonic functions for the most recent monthly index. There are also basic questions regarding which price data to use when estimating he- donic models. For instance, should data be collected on transaction prices or list prices? Although transaction prices seem preferable due to seasonal selling pat- terns, BLS has used regular list prices in hedonic modeling of apparel—the idea being to avoid looking at different points in a product’s life price cycle. The long list of unresolved issues discussed in this chapter explains why even some proponents of hedonics advocate a less aggressive expansion of its use in the CPI than BLS appears to be taking. It is important that the BLS position on hedonics be shaped by scientific corroboration of the validity of broadly applying the method across index items and not be adopted as the default method to correct for quality bias in an attempt to move the CPI closer toward a COLI ideal. There

EVOLVING MARKET BASKETS 145 is certainly no guarantee that hedonic methods always improve accuracy relative to alternative approaches. The data and specification problems discussed in this chapter are serious, and we believe that the value of hedonic methods, and of alternatives, must be determined over time on an item-by-item basis. This represents a major undertak- ing. Recommendation 4-7: Congress should continue to provide the BLS incremental resources to permit it to conduct in-depth and system- atic analysis of quality changes across a broad range of goods and services covered by the CPI. In designing its hedonics research program, BLS should seek to develop tools for dealing with the data and specification problems discussed above. Ex- tending the CPI improvement initiative will allow BLS to continue its experimen- tal research into scanner data; to assess the impact of hedonics on item compara- bility decisions and on index performance; and to investigate the replicability of competing techniques, perhaps using outside researchers to review and attempt to reproduce BLS results. Recommendation 4-8: An independent advisory panel, consisting of econometricians, statisticians, index experts, marketing special- ists, and possibly product engineers, should be formed to provide guidance on both conceptual and application issues pertaining to hedonic methods. BLS, working with the advisory panel, should assess the impact of modeling imperfections on the validity of its hedonic adjustments prior to their introduction into the index. This would provide an analytic basis for proceeding sensibly in the face of external pressures to ameliorate the perception that the CPI fails to capture improvements in rapidly evolving sectors and to proceed quickly in this area simply because it is viewed as the only option available. In addition to attempting to advance understanding of the econometric methodology underlying the esti- mation of hedonic functions, the proposed advisory panel should provide outside review to help guide decisions about potential new applications and about which BLS pilot studies are adequately developed to be incorporated into the index. The hedonic results should always be evaluated against BLS’s currently used alterna- tives (generally those associated with implicit quality adjustment techniques), as opposed to some idealized flawless solution. To improve its effectiveness, the proposed advisory panel might be charged with helping to promote a major academic research effort to address issues (like the validity of using brand-specific dummy variables in the regressions) that are suspect but are not currently being discussed in the literature. The initiative should aim to increase collaboration between BLS and outside researchers on

146 AT WHAT PRICE? both theoretical work and practical construction issues. The tendency to empha- size what can be most easily measured, rather than to focus on learning what characteristics are important to consumers, should be resisted. No research pro- gram can identify a universal set of criteria against which the BLS can validate its econometric procedures—there will always be a role for detailed case-by-case study. But precisely because so much judgment and knowledge of the product is involved, it makes sense to have outside review before new hedonic applications are brought into the CPI. TECHNICAL NOTE 1: BOSKIN COMMISSION ESTIMATES OF QUALITY CHANGE AND NEW GOODS BIAS In this note we briefly review the items, grouped into upper-level categories, that the Boskin commission identified as contributing significantly to its overall CPI bias estimate. We also make note of criticisms of commission methods by Moulton and Moses (1997) to illustrate the lack of consensus that exists regard- ing the magnitude of quality change and new goods biases—particularly at the level of disaggregated CPI component indexes. Food and Beverages The estimated bias associated with CPI pricing of fresh fruit and vegetables was the largest among components of the food and beverages category and was attributed by the commission primarily to the value to consumers of increased seasonal availability and variety. Limited by the dearth of published evidence on items in the food category, the commission was forced to lean heavily on Hausman’s (1997) work that calculated consumer surplus for a new variety of breakfast cereal as a means to quantitatively estimate the value consumers place on increased product variety. Citing data showing increased total consumption of products within the category, which they linked to increased variety and convenience, the commission arrived at an annual bias estimate of 0.6 percent for fresh fruit and vegetables. Moulton and Moses (1997) challenged this figure, showing that most of the increase in consumption over the period 1972- 1995 occurred after 1985, while most of the increase in availability occurred before 1985: “Part of the increase appears to have been driven by shifts in prefer- ences, perhaps as a response to improved knowledge about the health benefits of fresh vegetables” (Moulton and Moses, 1997:314). Shelter The Boskin commission produced detailed back-of-the-envelope calculations, based on assumptions about rental unit quality and size, to estimate a 0.25 percent annual bias for the shelter cost index. The commission’s position that CPI quality adjustments have been inadequate for shelter was deduced from the premise that newer apartments have increased significantly in quality (as reflected by improved amenities, such as central air conditioning) and in size (a

EVOLVING MARKET BASKETS 147 dimension of quality). They interpreted housing survey data as indicating that apartments increased in size by 20 percent between 1976 and 1993. Moulton and Moses (1997) countered, arguing that (1) rents generally do not increase propor- tionately with apartment size and (2) more importantly, that careful examination of data from the American Housing Survey and elsewhere suggests that the Boskin commission overstated historical increases in apartment sizes by perhaps a factor of three. Appliances and Electronics The commission’s bias estimates for this cat- egory are the largest—3.6 percent per year for the period 1973-1994 and 5.6 percent per year for 1994-1996. Due to the identifiable and quantifiable nature of appliance characteristics, and probably also to a priori notions about advances in the sector, research into this category of consumer spending is more extensive than for any other. Thus, the commission was able to access direct evidence, and the overall category estimate was extrapolated from items for which studies have been produced. The body of evidence included research by commission member Gordon (1990, cited in Boskin et al., 1996) of model-by-model comparisons from Consumer Reports. Moulton and Moses acknowledge that bias estimates for this category were probably the best documented by the Boskin commission: the report cites a number of academic and government studies that “develop hedonic adjustment models and find upward bias for personal computers, television, video equipment, and other items in this category” (Moulton and Moses, 1997:317). Apparel The Boskin commission used a “conservative reestimation” of figures from Gordon’s Sears catalog index, which rose less rapidly than the CPI subindex, to arrive at a 1 percent annual bias for the category. The main short- coming of the experiment, according to Moulton and Moses (1997), is that Gor- don measured year-to-year price changes only for the subset of apparel items that remained identical. The methodology links out—or deletes—the price increases associated with new product lines; the entire observed price change is assumed to reflect quality change. This approach produces misleading estimates if manufac- turers are most likely to hike prices when new lines and varieties are introduced, as suggested by BLS studies. Also, apparel prices are known to be affected by lower-level substitution bias because of cross-outlet and seasonal volatility that allows consumers to find similar items at very different prices, depending on the store and on shopping times. Because methods to minimize substitution bias have been applied by BLS to apparel items, Moulton and Moses (1997:318) note that “it is unclear whether the Advisory Commission avoided double counting when sorting through these various sources of bias to produce its estimate of quality bias.” Transportation (New and Used Vehicles/Motor Fuel) On the basis of studies showing increased quality and increased service lifetime, the Boskin commission estimated an annual bias of 0.59 percent for automobiles. The esti-

148 AT WHAT PRICE? mate was based on back-of-the-envelope calculations on the effect of increased longevity and, in turn, reduced depreciation rates, on annual operation costs. Triplett (1997), as well as Moulton and Moses, argues that the commission did not have accurate information about measures that BLS has implemented to take into account improved automobile quality. The Boskin commission also esti- mated a 0.25 percent annual bias associated with CPI pricing of motor fuel, which was attributed to failure of the CPI to capture convenience and time savings associated with automatic credit card readers at gas stations. Moulton and Moses offer their own back-of-the-envelope calculations, based on assumptions about the value of consumers’ time, time savings created by the machines, and average purchase size and find a bias about half as large. Medical Care The Boskin commission’s estimate of bias in the medical services index, 3.0 percent for both professional medical services and hospital and related services, is imputed largely from two empirical studies—Shapiro and Wilcox (1996) on treatment of cataracts and Cutler et al. (1996) on treatment for heart attacks. Thus, though Moulton and Moses agree that there is upward bias in the medical index, the validity of the commission’s estimate ultimately depends not only on the accuracy of these specific results but also on the extent to which the studied services are representative of the sector. Work by Berndt et al. (1996) and Griliches and Cockburn (1996) for prescription pharmaceuticals—for which the Boskin commission estimated a 2.0 percent per year bias—led BLS, in 1995, to change its method of pricing prescription drugs when generic versions become available. Also, beginning in January 1997, BLS adopted the PPI (Producer Price Index) method of pricing treatment-based bundles of hospital services. Both these measures reduced biases associated with measurement of medical service categories, although it likely did not eliminate them. Other Goods and Services The estimated biases associated with items other than those noted above were generally minor in terms of their impact on the all- item CPI. The Boskin commission suggested a 2.0 percent bias in sporting equip- ment and toys; small appliances such as hair dryers were assigned the same bias as large appliances, 3.0 percent per year. Personal financial services, a category for which output is extremely difficult to measure and rapid technological change (e.g., proliferation of ATMs and on-line account management) has occurred, the commission “conservatively” estimated an annual bias of 2.0 percent. The com- mission also discussed cellular phones but, as Moulton and Moses (1997:321) point out, it is not completely clear whether or not they included this in their estimated 1.0 percent bias for the “other utilities, including telephone” category.

EVOLVING MARKET BASKETS 149 TECHNICAL NOTE 2: MATHEMATICAL DESCRIPTION OF HEDONIC METHODS In the index number context, the hedonic function pi,t = ht(zi) for a product with multiple varieties—where pi,t is the price of the ith variety in period t and zi is a vector of the ith variety’s characteristics or attributes—plays the same con- ceptual role as the (scalar) price plays for an undifferentiated good. In the present context, the hedonic function can be viewed as a menu from which individual consumers make choices. A typical hedonic specification for econometric estimation uses the natural logarithm of an item’s price as the dependent variable and several characteristics as the explanatory variables. The model may contain discrete variables, indicat- ing whether or not a model has a feature, such as a CD drive on a computer, as well as continuous variables, such as the thread count of a fabric. Control vari- ables, such as purchase location or outlet type, may also be included. When, as is typically the case, the explanatory variables are included linearly (rather than, say, logarithmically), the coefficients can be interpreted as giving proportional changes in price associated with a one-unit change in the quality characteristic or from a switch in the dichotomous variable. If explanatory variables enter non- linearly, these proportional changes depend on the values of the explanatory variables. There is a large theoretical literature on the properties of observed hedonic functions (see, e.g., Rosen, 1974; Muelbauer, 1974; Feenstra, 1995; Barry et al., 1995; Diewert, 2001). Much of this literature is concerned with the extent to which ht provides information on producers’ costs and consumers’ preferences under various assumptions about the nature of competition. This is not our con- cern here: in general, hedonic functions are reduced-form reflections of details of tastes, technologies, endowments, and strategic behavior in differentiated prod- uct markets. In particular, when competition is imperfect, it is generally not possible to infer marginal costs from the observed hedonic functions. We follow most of the theoretical literature and assume what Pollak (1983) calls “Houthak- ker’s ‘heterogeneous’ or ‘H-characteristics’” approach, which fits products for which consumers purchase one and only one variety. (The alternative, “Lan- caster’s ‘linear and additive’ or ‘L-characteristics’” approach, applies when con- sumers purchase multiple varieties and care about the total amount of each char- acteristic supplied by all.) The use of hedonics in the index number context rests on being able to interpret the ht functions as summarizing the menu of alternatives faced by con- sumers in period t. This raises the general problem that different consumers in fact face different prices and have different stocks of information about their alternatives. Moreover, when price is not linear in the values of characteristics about which consumers care (see Muelbauer, 1974, for some relevant theory), which most hedonic studies seem to find, it follows that, even if ht is a smooth

150 AT WHAT PRICE? function, the marginal cost to consumers of any particular characteristic varies with z. Thus, consumers who choose different varieties of some product because of differences in incomes or tastes (or both) face different “characteristics prices” at the margin, and the “characteristics prices” faced by nonbuyers are clearly not well defined. Finally, most uses of hedonics involve using an estimate of ht to, in effect, predict the price that would have prevailed in period t for a variety or model not actually offered for sale in that period. While this seems sensible, it is problem- atic at the theoretical level: under imperfect competition, if an additional variety or model had actually been offered for sale, the prices of other products might also have changed. In addition, smoothness and functional form assumptions are important in these exercises, and, particularly when consumers are heteroge- neous, theory provides relatively little guidance regarding such assumptions (see Diewert, 2001, for a useful discussion). The Indirect Method As discussed in the body of the chapter, the indirect method is used to handle situations in which one variety of a good tracked in the CPI system—with a specific vector of characteristics z1 and price p1,t , say—disappears after period t and is replaced by another—with characteristics z2 and price p2,t+1, beginning in period t + 1. There are two basic types of indirect methods. If the hedonic function, ht(z), for period t is available, the simplest form of the forward-looking indirect method involves using p2,t+1/ht(z2) as the estimated “pure” price relative. The denominator of this ratio is an estimate of what a good with characteristics z2 would have cost if it had been available in period t, based on the empirical relation between price and characteristics in that period. If the hedonic function, ht+1(z), for period t + 1 is available, the simplest backward-looking indirect method involves using ht+1(z1)/p1t as the estimated price relative. Because it uses a bundle (of characteristics) purchased in period t + 1, the forward-looking method is Paasche-like; similarly, the backward-looking method is Laspeyres-like. Hedonic functions are typically refit only periodically, so neither the current period nor the prior period function is usually available. Thus, the backward- looking method is rarely feasible. To see how this affects the calculations under the forward-looking method, suppose the hedonic function was last estimated in period 0, with h0(z) the estimated function, and suppose one wants to calculate the “pure” price relative between periods t and t + 1. Clearly, p2,t+1/h0(z2) is a forward-looking estimate of the price relative between periods 0 and t + 1 for the bundle z2, while p1,t/h0(z1) gives a similar estimate of the price relative between periods 0 and t for bundle z1. If z1 and z2 were the same bundle, the ratio of these quantities Rt ,t +1 = [ p2,t +1 / p1,t ][h0 ( z1 ) / h0 ( z2 )] , (1)

EVOLVING MARKET BASKETS 151 would give the price relative between periods 1 and 2 for that bundle. Since product 2 is being treated as a replacement for product 1, z1 and z2 must be close in some relevant sense. In any case, the BLS proceeds as if they were equal and employs Rt,t+1 as the price relative. Another way to look at (1) is that the actual price of product 2 in period t + 1 is being compared with the adjusted price of product 1 in period t—adjusted for the quality difference between products 1 and 2 using h0(z): adjusted p1,t = p1, t [h0 ( z2 ) / h0 ( z1 )] . (2) It is easy to show that these methods automatically take into account some forms of unobservable outlet-specific price differences that, along with other factors, prevent hedonic functions from fitting perfectly. Suppose, for instance, that h0(z) is the estimated marketwide base period hedonic function, as above, but prices of all varieties in some particular outlet exceed marketwide averages by a constant multiple θ. Then p2/θh0(z2) is the natural estimate of the price relative between periods 0 and 2 for the bundle z2, while p1/θh0(z1) is the natural estimate of the price relative between periods 0 and 1 for bundle z1. Neither of these is observable if θ is unknown, but their ratio, which is the quantity of interest, is given simply by equation 1, above. The Direct Time Dummy Method This method involves estimating hedonic functions of the following form: T log( pi,t ) = h( zi,t ) + ∑ βτ δ (t, τ ), for t = 0, . . . , T ; all i , (3) τ =1 where the subscript i denotes varieties or models and, as above, the βτ are con- stants, and δ(t,τ) equals 1, if t = τ and 0 otherwise. Note that there is no time dummy for period 0, the base period; we have arbitrarily normalized at β0 = 0 to identify the rest of the model. Specification (2) implies that in any period t the ratio of the prices of models with, say, characteristics bundles z1 and z2, p1,t/p2,t, is equal to antilog[h(z1) – h(z2)], which does not depend on time. It is thus being assumed that the prices of all (actual and potential) varieties change proportionately over time. (In light of Zvi Griliches’s seminal contributions to the theory and practice of hedonic meth- ods, the panel believes it would be appropriate to label this the case of Griliches neutrality.) Neither theory nor empirical research provides much support for this assumption, however, particularly in industries experiencing rapid technological change. If prices of all varieties do change proportionally, though, it is simple to use the function above to produce a “pure” price relative for the product under study. For any variety i with a constant characteristic vector zi, the equation above immediately implies that for any two time periods t and u

152 AT WHAT PRICE? pit / piu = antilog[log( pit ) − log( piu )] = antilog[β t − β u ] , (4) for all i. Thus the expression on the right gives the price relative between periods u and t for the product under study. The Direct Characteristics Method Let ht(z) be the hedonic function in period t, Ct be the set of varieties avail- able—with characteristic vectors zi,t, (average) prices pi,t, and quantities sold qi,t. The direct characteristics method computes price relatives using these data with- out necessarily imposing the assumption (which underlies the time dummy method) that ratios of hedonic functions are independent of the point in character- istics space at which they are evaluated. If some coefficients of the hedonic function are constant over time, of course, estimation efficiency can be improved by imposing constancy and using data from multiple periods in estimation. Alter- natively, if the assumption that all slope coefficients are stable over time (i.e., the assumption of Griliches neutrality that underlies the time dummy method) is rejected by statistical test, some use of some version of the direct characteristics method would seem to be in order. As noted in the text, the natural way to use the hedonic functions to compute a single price relative in, say, periods 1 and 2, with different sets of products available in each, is to use the hedonic functions to price constant bundles of characteristics over time. The literature suggests two ways of doing this. The first follows Diewert (2001) and uses the average bundles consumed as reference characteristics vectors: zt∗ = Σ i qit zit / Σ i qit , t = 1, 2 . (5) Then Laspeyres-, Paasche-, and Fisher-type indexes, which give alternative mea- sures of price relative between periods 1 and 2, can be defined, respectively, as follows: L12 = h2 ( z1∗ ) / h1 ( z1∗ ) , (6a) P12 = h2 ( z2∗ ) / h1 ( z2∗ ) , (6b) F12 = [ L12 P12 ] . 1/ 2 (6c) Note that (6a) requires only lagged quantity weights, while both (6b) and (6c) require current quantity data. Note also that all these measures are equal, and all equal the results of the time dummy method, if the ratio h2(z)/h1(z) is independent of z. The second approach follows Feenstra (1995), with some modifications by Diewert (2001). Let C* be the set of varieties available in both periods, and let Ct′ be the set of varieties that are available only in period t. One can use the period 1 hedonic function to “predict” the period 1 prices of those varieties available only in period 2, and one can use the period 2 hedonic function similarly:

EVOLVING MARKET BASKETS 153 pi1′ = h1 ( zi 2 ), zi 2 ∈ C2 ′ , (7a) pi 2 ′ = h2 ( zi1 ), zi1 ∈ C1′ . (7b) One can then compute a Laspeyres-like measure by taking a weighted average, using period 1 sales shares as weights, of the (actual and “predicted”) price ratios of the varieties available in period 1: L12 = ΣC* wi(pi2/pi1) + ΣC1′ wi(pi2′/pi1) = (8a) [ΣC* qi1pi2 + ΣC1′ qi1pi2′]/ΣC1 qi1pi1, where, as usual, wi = qi1pi1/ΣC1 qi1pi1. Similarly, using period 2 sales shares of the various varieties as weights and “predicting” the period 1 prices of varieties available only in period 2 yields a Paasche-like measure: P12 = ΣC* wi(pi2/pi1) + ΣC2′ wi(pi2/pi1′) = (8b) [ΣC2 qi2pi2]/[ΣC* qi2pi1 + ΣC2′ qi2pi1′], where wi = qi2pi1/[ΣC* qi2pi1 + ΣC2′ qi2pi1′] for zi2∈C*, and wi = qi2pi1′/[SC* qi2pi1 + SC2′ qi2pi1′] for zi2ŒC2′. One can combine these, as in (6c), to obtain a Fisher- like measure of the price relative for this product. Note again that if price ratios for all varieties are the same, as assumed by the time dummy method, all of these measures are equal. To see the sense in which these two approaches give Laspeyres-like and Paasche-like measures, it is instructive to follow Pakes (2001) and consider a single consumer with income y in periods 1 and 2, with prices the same in both periods for all goods but widgets. In period 1, the consumer has available a set of varieties C1, the prices of which are given by the known hedonic function h1(z), and she purchases variety z1. In period 2, the consumer faces choice set C2 and known hedonic function h2(z), and she chooses variety z2. Suppose this consumer is given h2(z1) – h1(z1) additional income in period 2. Is this greater or less than the compensating variation, the period 2 income increase that would leave her exactly as well off as in period 1? If z1∈C2, buying variety z1 in period 2 would leave her with y + h2(z1) – h1(z1) – h2(z1) = y – h1(z1) to spend on other goods, exactly as in period 1. So h2(z1) – h1(z1) is at least equal to the compensating variation. But because the two hedonic functions are differ- ent, it may be possible for the consumer to do even better by choosing some z2′ π z1 in C2. Thus h2(z1) – h1(z1) is greater than or equal to the compensating varia- tion, depending on whether such a z1′ exists or not. Similarly, suppose instead that the consumer’s period 1 income is reduced by h2(x2) – h1(x2). Is this greater or less than the equivalent variation, the period 1 income reduction that would leave her exactly as well off as in period 2? If z2∈C1, buying variety z2 in period 1 would leave her with y + h1(z2) – h2(z2) – h1(z2) = y – h2(z2) to spend on other goods, exactly as in period 1. So this income

154 AT WHAT PRICE? reduction will leave the consumer no worse off. She will be better off in the (new) first period if she can afford some z1′∈C1 that she prefers to z2. Thus h2(z2) – h1(z2) is less than or equal to the equivalent variation, depending on whether such a z1′ exists or not. For our single consumer, the price relative could naturally be computed as either h2(z1)/h1(z1) or h2(z2)/h1(z2). The former is a Laspeyres approach and, as above, relates to the compensating variation. The latter is a Paasche approach and relates to the equivalent variation. In the usual sense, and with all the usual caveats plus the requirement that z1 and z2 be available in both periods, in this simple case these two measures bound the true, preference-dependent, change in the cost of living.

Next: 5 New Goods and New Outlets »
At What Price?: Conceptualizing and Measuring Cost-of-Living and Price Indexes Get This Book
×
Buy Hardback | $54.95 Buy Ebook | $43.99
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

How well does the consumer price index (CPI) reflect the changes that people actually face in living costs—from apples to computers to health care? Given how it is used, is it desirable to construct the CPI as a cost-of-living index (COLI)? With what level of accuracy is it possible to construct a single index that represents changes in the living costs of the nation's diverse population?

At What Price? examines the foundations for consumer price indexes, comparing the conceptual and practical strengths, weaknesses, and limitations of traditional "fixed basket" and COLI approaches. The book delves into a range of complex issues, from how to deal with the changing quality of goods and services, including difficult-to-define medical services, to how to weight the expenditure patterns of different consumers. It sorts through the key attributes and underlying assumptions that define each index type in order to answer the question: Should a COLI framework be used in constructing the U.S. CPI?

In answering this question, the book makes recommendations as to how the Bureau of Labor Statistics can continue to improve the accuracy and relevance of the CPI. With conclusions that could affect the amount of your next pay raise, At What Price? is important to everyone, and a must-read for policy makers, researchers, and employers.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!