New Goods and New Outlets
This chapter is essentially a continuation of the quality change discussion. In the first half of the chapter we consider the case of new goods that do not fall into existing Consumer Price Index (CPI) item categories. This case presents problems of estimating newly created value, sample rotation frequency, item reclassification, and weight updating. The second half of the chapter addresses the indexing problem that arises when consumer shopping patterns are shifting. The panel considers what, if anything, the Bureau of Labor Statistics (BLS) could do to identify and estimate quality and price components of observed differences in the prices of goods across outlets.
The term “new good” is not precise. Routine price collection procedures continually lead to instances in which BLS has to find replacements for items that have disappeared. Similarly, when BLS rotates its sample of retail outlets, it picks up products that are different from those it had been pricing in the old stores.1 Products also appear that are novel to the point that there is no place in the CPI
item structure to accommodate them: cell phones, home computers, and VCRs are examples.2 These are products whose characteristics would be difficult to “repackage” (in the sense discussed in Chapter 4) into existing goods and services no matter how broadly definitions are drawn. Without an explicit decision to change the list of goods to be priced, standard indexing procedures will not pick up any of the effect of such newly introduced items on consumers’ living standards or costs.3
These contrasts notwithstanding, no sharp dividing line separates a new good from a quality improved product. What can be cleanly distinguished are situations that lead to within-sample item replacement and those involving a good or service that has entered the market but would never be brought into the index as part of the in-store pricing process. New items falling into this second category include (1) those that might be picked up during sample rotation (in which case items enter using overlap pricing, where there is no comparison to a previously priced good and, hence, no quality adjustment) and (2) those that can only be brought into the index when item strata are redefined and the sample reset. The previous chapter primarily addressed shortcomings in the process for dealing with quality adjustment of replacement items. However, failure to capture price (or cost-of-living) effects associated with new nonreplacement products may, depending on the objective of the index, cause what Triplett (2001b) has termed “new introduction bias.” This failure is not a quality adjustment problem but a sampling one—a case in which rapid product turnover, caused by technological or other changes, leaves the item sample no longer representative of what people buy (Triplett, 2001b:19-20).
The appearance of products that can only enter the index after item reclassification (and, to some extent, those that enter during sample rotation) raises two issues beyond those associated with routine item replacement. The first is what to do to account for price effects that occur during the period in which a new product appears on the market. Specifically, should a price or cost-of-living index reflect the fact that new goods typically enter the market at a price that is below that which would have reduced demand in the period prior to its introduction to
zero or, conversely, for an item that has disappeared, the increase in price that would have driven quantity demanded to zero? Failure to capture this “price reduction” could be argued to cause a new good introduction bias in the index. The second issue is what to do in subsequent periods after the appearance of the new good. Specifically, how and when should these items be brought into the market basket tracked by the index?
Valuation at the Point of Introduction
As new products penetrate the market, the item coverage of a fixed-basket index becomes less and less representative of the things that people are currently buying. This is why various techniques of “unfixing” the market basket, including item replacement and sample rotation, are now regular features of the CPI. However, even with the modifications that these techniques allow, potential indexing problems remain. Identifying one of these problems, Hausman (1997:209) argues that welfare effects associated with the introduction of new goods should be estimated and used to adjust the CPI:
The CPI serves as an approximation of an ideal cost-of-living (COL) index. In turn, the COL index answers the question of how much more (or less) income a consumer requires to be as well-off in period 1 as in period 0 given changes in prices, changes in the quality of goods, and the introduction of new goods. . . . The CPI as currently estimated by the Bureau of Labor Statistics (BLS) does a reasonable job of accounting for price changes and has begun to attempt to include quality changes. However, the BLS has not attempted to estimate the effect of the introduction of new goods, despite the recognition of the potential importance of new goods on both a COL index and the CPI.
In this subsection we briefly review the mechanics of how, in theory, price indexes could take into account this new goods effect. We then consider a counterargument to Hausman’s (and, implicitly, the Boskin commission’s) recommendation for BLS to do so.
The relevant difference between a new good and an established one is that, for the former, the price in previous periods associated with the realized sales levels (zero) cannot be observed, while the price for the established good can be. In theory, a virtual price exists in each prior period that would have been just high enough to drive the quantity demanded of the new good to zero.4 There are consumers who would have purchased the good at various prices between that virtual price and the lower price at which the good sells when it appears in
markets.5 The effective decrease from the virtual to the introductory price of a new good is not captured in the CPI, even in instances when new goods are brought into the market basket very rapidly. The introduction of a new good, and its later diffusion to its ultimate customer base as more consumers learn about it, may be thought of as a series of price reductions. A demand curve that traces the “virtual” prices that some consumers would have been willing to pay for the good can, in theory, be econometrically estimated.6
If significant numbers of new goods are continually invented and successfully marketed, an upward bias will be imparted to the overall price index, relative to an unqualified COLI (though this effect may be partially offset by a downward bias created by the disappearance of goods). There is a component of this bias that can occur even if new goods are linked into the index quickly and expenditure weights are updated frequently. A priori, one might expect that only new goods that provide radically improved capabilities or that are sold at reduced prices relative to predecessors would capture market share quickly enough to generate significant point-of-introduction bias. After all, if the new good offers only minor new capabilities relative to existing goods, the virtual price driving
consumer demand for it to zero would not be far above that for close substitutes.7 Conceivably, BLS could attempt to identify and estimate demand curves for radically different goods experiencing rapidly growing consumer acceptance. The price or cost-of-living index would include weighted estimates of the difference between the virtual price (in period t - 1) and the introductory price (in period t). To be fully consistent, it would also be necessary to identify goods that were forced off the market by newly introduced competitors, the unavailability of which created virtual price increases.
Hausman (1997) argued that it is important to consider new brand introduction in the calculation of economic welfare and consumer price indexes and then proceeded to estimate the demand curve for a new variety of breakfast cereal— Apple-Cinnamon Cheerios—to illustrate that it could be done. His article concludes that introduction of the new cereal variety—hardly a radical expansion of consumer opportunities—created substantial consumer benefits (defined by the difference between the virtual price for the cereal minus the introductory price), sufficient on its own to have reduced the average rate of price increase in the CPI component for breakfast cereals significantly (Hausman, 1997:229, 234):
. . . to the extent that about 25 percent of cereal demand was from new brands over the past ten years, and under the (perhaps unrealistic) assumptions that the new brands sell for about the same average price as existing brands and that the estimate here would generalize to a reservation price of about two times the actual price, the overall price index for cereals which excludes the effects of new brands would be too high by about the overall share of new brands—25 percent. . . . The introduction of imperfect competition would reduce the overstatement of the cereal CPI to about 20 percent.
If this kind of differentiation produces a significant number of products that enter the market priced well below that which would choke quantity demanded to zero, then our economy, marked as it is by increasing proliferation of product varieties, must be producing a substantial stream of new consumer welfare. This would strengthen the case for making price imputations to account for “price” effects attributable to the introduction of new commodities to the marketplace.
On the other hand, there are weaknesses in the case. Hicks (1940) defined market demand as it relates to new goods, and Hausman demonstrated that a choke price could be estimated for a specific new good. However, there is no clearly acceptable technique for consistently estimating demand curves for new goods or services in such a way that choke prices can be confidently ascertained. Several panel members are also, independent of estimation problems, hesitant
about the advisability of doing so on conceptual grounds (these concerns are discussed in the final section of this chapter).
The practical problem facing statistical agencies is exactly how the virtual price of a new commodity should be estimated. Many aspects of Hausman’s analysis are highly controversial, even in the context of microeconomic research that is not directly tied to policy. Hausman’s findings have been disputed on the grounds that questionable assumptions were built into his econometric specifications, which led to substantial overstatement of the prices that consumers would have paid for the new cereal brand. In a response to Hausman (1997), Bresnahan (1997) argued that Hausman’s model requirements—specifically that there be no demand shocks that cause consumers to shift purchases and that shocks are not reflected in prices because they are unanticipated—are inappropriate. The assumptions required about functional form (the shape of the demand curve) and for system identification simply introduce too much uncertainty to be used as an input into a statistic that must be produced in a replicable fashion on a regular basis. Calculating the price that drives demand for a product to zero requires extrapolation outside the range of price and quantity observations. Hence, it must rely more heavily than conventional demand estimation on untestable assumptions about functional form. Bresnahan (1997:237, 246) concluded that the question of how important new goods are in terms of their contribution to social welfare remains unanswered.8 Procedures for estimating virtual prices would require extensive refinement before they could even be considered for adoption into a national price or cost-of-living index.
Research into welfare and price effects associated with new goods is important and deserves attention, but it is unlikely that such a program will produce a consensus methodology in the near future. Given the level of uncertainty among economists about the accuracy and replicability of current econometric techniques for estimating virtual demand, it would be imprudent for BLS to attempt to adjust the CPI to account for increased welfare that occurs at the point when new products are introduced.9
Conclusion 5-1: Virtual price reductions associated with the introduction of new goods should not be imputed for use in the CPI.
Several members of the panel—particularly those advocating separate price and cost-of-living indexes—are unconvinced that adjusting the CPI to account
for point-of-introduction bias would be a good idea even if the practical estimation problems could be solved. Proponents of more traditional price index methodologies argue that it is a perversion of the language to argue that the effect of, say, the introduction of cell phones or the birth control pill is to reduce the price level, a result that comes from confusing the concept of a price level with that of the cost of living. Their position is tempered somewhat by the realization that, outside of price measurement, there is nowhere else in the national accounts for such product quality improvements to be included and, as Nordhaus (1998) and others have argued, real growth in the economy is thereby understated. Additionally, modern economic growth appears to be more quality intensive than quantity intensive, and the statistical system is not keeping up with the change. However, the panel as a whole agrees that adjusting the CPI is not the way to correct the situation. Rather, research in this area should be directed toward developing a separate experimental COLI that is adjusted, to the extent possible, to account for changes as new products and technologies diffuse throughout the economy.
Criteria for Introducing New Goods
Under traditional procedures, a new (nonreplacement) good is linked into the CPI in such a way that its introduction, in and of itself, has no effect on the level of the index. Once in the index, price change of the new item affects index growth in the normal fashion. However, in addition to the point-of-introduction price reductions discussed above, price trends over the interim period between product appearance and introduction to the index also go uncaptured. Thus, a second problem—that of how quickly new goods are brought into the index—exists if early price-cycle patterns are consistently different from general price trends. If a new commodity is a reasonably close substitute for an existing one and is likely to replace it in the marketplace, then instead of explicitly revising the base market basket, one could think of simply replacing the old commodity in the index with the new commodity, after some adjustment for quality change. This is not very different from the within-sample replacement that occurs when an outlet sample item disappears.
For more novel introductions, a new commodity must be brought into the index as part of a revision to the market basket; that is, when the statistical agency switches from the old fixed-basket Laspeyres index to a new fixed-basket Laspeyres index that has a more recent period as its base and includes the commodity. There may still be a problem with use of the latter in comparison with a superlative index because of the properties of new product price cycles. A Laspeyres index that has period 1 as its base will weight the long-term price relative for the new good by its period 1 market share, which will often be much smaller than its period t market share for t > 1. Note that this period t market share
appears for the period t Paasche index. Thus, under the above conditions, a Paasche index will often be considerably lower than its Laspeyres counterpart.10
Typically, at least in high-tech sectors, a new good does come to the market at a relatively high price and initially has a small volume of sales. High prices may reflect both production costs that have yet to be reduced by learning and process innovation and seller attempts to maximize profits by first selling to customers for whom the new product is especially valuable. Subsequently, prices often fall.
Armknecht et al. (1997) describe the price cycle observed after the introduction of VCRs to the market. Early in the product cycle, in 1978, approximately 400,000 units were sold at an average price of more than $1200. By 1987, when VCRs were introduced into the CPI, annual sales had increased to almost 12 million units, and the average price was less than $500. Over this period, prices decreased by 60 percent, non-quality adjusted, and, by the time of introduction to the index, sales accounted for about 0.1 percent of consumer expenditures (Armknecht et al., 1997:388). Dulberger (1993) showed how, for such products, more frequent replenishing of item samples can have a large effect on measured price change. Her analysis of semiconductors produced a chained Fisher index that decreased by 29.2 percent per year when new chips were introduced into the index with only a 1-year delay after appearing on the market. However, the index only decreased by 20.1 percent when there was a 3-year delay, and the index showed virtually no price decrease when the lag was 5 years (Dulberger, 1993:Tables 3.7, 3.8).
Observed price patterns such as these have led to charges, by the Boskin commission and others, that delayed introduction of new goods systematically omits product-cycle dynamics that impart an upward bias on price indexes. More frequent updating of the item classification structure and of the sample (which, in turn, would require more frequent index chaining) would have allowed a greater portion of these early product price trends to be captured and led to a more accurate plutocratic index. Each case, individually, would not have had a large cumulative effect on the overall CPI (for VCRs, well under one-tenth of 1 percent). However, in the modern economy, a large number of new goods are introduced each year, each having some effect. It is important to note, though, that not all goods follow this kind of pricing path during their life cycle. Pakes (1997), for instance, has stressed producer efforts to penetrate markets with low introductory
As discussed in Chapter 2, the Paasche and Laspeyres indexes are both valid measures of price change between periods. If only one measure of overall price change is required, it can be argued that an average of these two indexes, such as the Fisher ideal index, which takes the geometric mean of the two indexes, offers a sensible approach. Under the conditions outlined, the Paasche and Fisher indexes will give a lower measure of price change. Thus, one is again led to a strong argument for the production of a superlative index in addition to the present real-time CPI.
prices.11 Since pricing early in goods’ life cycles may be atypical in different ways, the effects of the loss of information on pricing during an inevitable period of delay are, at least in principle, indeterminate.12
At a cost, criticisms by the Boskin commission and others could certainly be addressed. New goods could be introduced into the index earlier so as to catch a larger portion of postintroduction price trends; growing sales can be captured by more frequent weight updating. This panel agrees that, other things being equal, more information is better than less. Hence, new commodities should in principle be introduced into the CPI as soon as they become significant in the marketplace. Such an approach would require frequent sample rotation to capture new supplemental goods and more frequent revision of the item classification structure to capture radically different goods.
Unfortunately, other things would not be equal. Survey-based updates of expenditure weights and product samples are expensive to produce and require time to compile. When a new commodity is introduced into the base without new household expenditure information, other complications arise: Should all the old expenditures simply be scaled down proportionally to make room for the new expenditure share? Or should the expenditure shares of commodities that are apparently the closest substitutes for the new commodity be reduced somehow to make room for the new expenditure share? Again, this depends in part on how the new good enters the index.
As noted above, some items, such as VCRs, only enter during item reclassification or, like Viagra, from an ad hoc targeted initiation. However, the majority of supplemental and new items are identified in the Telephone Point of Purchase Survey (TPOPS, the 1998 revision of the POPS) and enter the CPI during sample
rotation. New item and outlet samples are drawn annually for a subset of the 218 defined TPOPS item categories, so the entire sample turns over periodically (the previous procedure rotated geographic areas rather than item categories). Because of the importance of this process in getting new products into the CPI basket, the Conference Board (1999:25) recommended that BLS speed up sample rotation. Citing how quickly new products enter the marketplace in the modern economy, the group recommended an eventual 2-year rotation schedule for most categories and an annual rotation for categories subject to frequent change.
It is important to point out that the BLS has made significant strides in improving its survey structure to decrease time between outlet and item rotation. Part of the CPI Improvement Initiative was used to begin data collection procedures designed to incorporate new goods into the index more quickly. In addition to moving from an area-based to an item-based outlet rotation process, TPOPS has shortened the amount of time necessary to draw an outlet sample. Instead of one-fifth, one-fourth of outlet samples (and contemporaneously item samples) will soon be rotated each year, which decreases the amount of time needed for full rotation from 5 years to 4. Lane (2000:8) notes that, “by 2003, when the CPI has initiated an entire cycle of outlets based on TPOPS, the outlet samples will be significantly more current than they were before 1999.” More frequent rotation might also be complemented to some degree by expanded use of targeted and directed replacement procedures with which BLS is currently experimenting. Targeted outlet item rotation would allow TPOPS categories associated with quickly changing markets to be rotated on a fast-track basis; targeted item rotation involves increasing, ad hoc, the probability of selecting specific items. BLS is also looking at methods for decoupling item and outlet rotation so that items could be rotated (within the current sample of outlets) without waiting for outlet rotation. Targeted replacement is suggested since outlet rotation is a particularly time-consuming and expensive aspect of CPI sampling. Rotation only at the item level may offer a way around this practical constraint, at least for items that enter the index through stores (or types of stores) already represented in the sample.
Sampling is done for 300-400 entry-level items (ELIs), like oranges, which are more finely specified examples of the 218-item strata, such as citrus fruits. Item strata correspond roughly to the TPOPS categories. If new items encountered in the sample rotation process fit existing ELI definitions, they can be readily brought into the CPI through overlap pricing, since new and old items are both available in at least one period (in which case the base period price difference between the new and old items is implicitly treated as completely due to quality differences).
From time to time, new items appear in the system that do not fit existing ELI definitions but are close substitutes for items that do. The example cited in Armknecht et al. (1997) is CD players, which were clearly substitutes for phonographs and tape players. Such items emerge in Consumer Expenditure Survey (CEX) interviews and are coded separately so that expenditure data can be en-
tered; Armknecht et al. (1997) indicate that existing CPI procedures can accommodate these situations. However, “because of the time lapse associated with the CEX and TPOPS surveys, unless special efforts are made to reinitiate an item stratum, three to four years elapse from the time the product appears in the marketplace to the time the product appears in the CPI” (Armknecht et al., 1997:387). BLS accounts of the 1998 revised CPI methodology, however, suggest that the new ELI structure and the change to category rotation under TPOPS will provide more sampling flexibility and bring new items more quickly into the index (see Greenlees and Mason, 1996).
If costs linked to these processes were not a consideration, the panel would likely agree, without qualification, with the Conference Board’s suggestion that sample rotation schedules should be further sped up for most goods. Cost is a big factor though and, given current evidence, it not clear that a 2-year rotation plan would yield commensurate benefits in terms of index accuracy.
Recommendation 5-1: Until it can be shown that further compression of the sample rotation schedule would create significantly different rates of change in the CPI, the panel is satisfied with the current BLS plans. We do recommend that BLS undertake research designed to assess the impact that moving from a 4-year to a 2-year rotation cycle would have on the rate of index growth.
This will entail analyzing a broad set of items to simulate the effect that more frequent rotation would have. BLS should specifically investigate the statistical disadvantages of frequent sample rotation. New samples enter the CPI by chaining, and these transient disturbances can cause an index to display higher variances and a tendency to drift upward. Only after these issues are given further attention can recommendations regarding the optimal sample rotation frequency be advanced in a fully informed manner.
As noted above, BLS must also deal with the class of product introductions that are not picked up during normal sample rotation. If a new item is encountered that does not fit an existing ELI definition and is not obviously a close substitute for one that does but does fit within an existing item stratum, the remedy is to define a new ELI. “This process could take five to seven years for full implementation” through the normal sample rotation process (Armknecht et al., 1997:387).
If a new item does not fit any existing item stratum definition (e.g., cell phones, VCRs), it normally does not enter the CPI until item strata are redefined. Until recently, this happened only when major revisions of the CPI were introduced—about every 10 years.13 Historically, the more novel a new item, the
longer it has taken to appear in the CPI. Home computers and VCRs, for instance, were not introduced into the CPI classification until 1987.
BLS now recognizes that the past delays between the introduction of entirely new and important goods and their appearance in the CPI are no longer acceptable. The most obvious way that this recognition has changed the CPI production schedule is in the updating of upper-level expenditure weights. Weights will be updated more frequently, and they will be based on a shorter span of expenditure data. Beginning in 2002, weights will be updated every 2 years, with a 2-year lag (Bureau of Labor Statistics, 1998a):
Thus, for example, CPI expenditure weights will be updated to the 2001-02 period effective with release of CPI data for January 2004. As a result of this change, expenditure weight data will be, on average, “two years old” when introduced into the CPI, and four years old when replaced. By contrast, the most recent set of CPI expenditure weights—based on 1993-95 CEX data— were on average, 3 years old when first used in January 1998, and they replaced weights that were about 15 years old.
BLS has also begun researching the advisability of adopting special targeted procedures to quickly bring new products (Viagra was one) into the index (for a full description, see Lane, 2000). BLS’s work in this area is commendable, and BLS should continue to develop changes in its procedures designed to reduce those delays substantially.14 When visible items, such as home computers and VCRs, that achieve significant expenditure shares can be brought into the item samples rapidly, public and policy maker confidence in the CPI and the BLS can only be improved. While the effect of earlier inclusion of any one product is
likely to be minor, systematic early inclusion of new goods broadly could have a significant effect on index growth.
There are limits on the extent to which speeding up product introduction will relieve the problem, however, since the BLS may not be able to justify—for both analytic consistency and budgetary reasons—revolving a large number of new goods on an extremely short-cycle schedule. Analytically, more rapid rotation and more frequent rebasing require proportionately more chaining of indexes with non-identical components, which can exacerbate index drift (see Glossary). On the cost side, each rotation creates inefficiency because another period must elapse (to produce a price change) before quotes on new items can be used. As rotation frequency increases, the amount of information used in the index relative to the total amount of information collected is decreased.
Finally, though it is important that new items be introduced into the index once they are commonly consumed, they need to be entered with a correct expenditure weight. Since only price data are compiled on a monthly basis, it is not easy to estimate a weight immediately. Thus, there may be a tradeoff between timeliness and accuracy of item weights. For many cases, it may only be practical to introduce new items into the sample rotation after a significant and estimable market share emerges.
In light of these considerations, two approaches to the item introduction problem seem potentially worth considering by the BLS. First, broader ELI and item strata definitions or definitions couched in terms of function instead of product (e.g., audio reproduction instead of phonographs and tape players) might reduce substantially the number of new items that must be excluded for long periods because they do not fit existing definitions.15 We recognize that there is likely a tradeoff between breadth and clarity of boundaries, but that does not establish that more breadth would not be better. Second, the Conference Board suggests expanding the BLS’s small program of special sampling to “arrange for regular consultations with panels of experts . . . persons who are likely to know when important new consumer products have recently or soon will reach the market.” This, too, seems sensible and worth serious consideration.
BLS rotates a portion of its sample of retail stores and business establishments each year. The probability of an outlet being selected is proportional to store-by-store expenditures reported by consumers in TPOPS. Outlet rotation is
designed to make the index reflect changing consumer shopping patterns—that is, to collect prices from the places where consumers shop most. Item rotation and outlet rotation occur simultaneously; as a new outlet rotates into the sample, so too do many new (to the index) products. A new product is introduced to cover the same market basket category (for instance “records and tapes”) as that which was sampled at the outgoing outlet, but whether or not the same specific item type (say, a CD) that was previously priced remains in the sample depends on expenditure shares at the new store in conjunction with the random selection component of the rotation procedure.
In the previous section we addressed difficulties that arise when sales-based outlet and item selection require BLS to compare newly sampled, non-identical items with those that have been replaced from the outgoing outlets. However, a different issue arises when, after outlet rotation, a specific product from the old outlet sample also appears in the new outlet sample but at a different price. Because of the expenditure-based sampling process, products with high market share and high sales volume are particularly likely to be reselected through successive outlet rotations.16
Outlet substitution expands the scope of index coverage beyond that represented by specific items, incorporating the notion that, in acquiring goods and services, aspects other than price may affect a consumer’s cost of living.17 When consumers purchase a good at a particular store they are buying a package. The package includes not only the specific item but also the quality of the shopping experience—the services provided, the store’s locational convenience, its return and exchange policy, and the variety of products available. In this context, the issue of the value of time naturally arises. As it relates to consumer shopping, and specifically outlet-use patterns, valuing nonmonetary benefits associated with time savings, improved convenience, or better service is central to the concept of a COLI. Treatment of the issue should distinguish between time as (1) a variable that (perhaps combined with transportation costs) might be used to help explain differential outlet quality and (2) any explicit imputations of the value of time spent shopping (perhaps corrected for any entertainment component of the activ-
ity), derived from wage, survey, or other data, which might be added to the cost of obtaining a given level of material well-being.
Variation in prices charged for the same goods at different kinds of outlets can be substantial. For example, it is not uncommon for a specific brand and size of some good—a laundry detergent or motor oil—to cost 50 to 100 percent more at a 24-hour convenience store than at a large discount outlet. As Pollak (1998) notes, consumers face a distribution of prices for many goods and services, rather than a single price, and at least some of them find it worthwhile to search for lower prices.18 Consumers benefit if they can reduce costs—and if these gains are not fully offset by inferior service or greater item acquisition costs—by substituting purchases from a high-price seller to a low-price competitor. But because of the way in which new samples are linked to old ones, this type of consumer benefit is not always picked up correctly by CPI procedures.
When a new set of stores enter the sample and prices for various categories of goods are collected, all of the difference between the old outlet price and the new outlet price is linked out. For cases in which an identical item (such as a 12-ounce tube of Crest toothpaste) is priced and that price is different, BLS attributes the entire difference to outlet-related quality variation. This could be correct in some cases; however, for all cases, any change in the price recorded for the item has no effect on the CPI.19 Application of this procedure implies acceptance of the assumption that markets are in equilibrium, so that differences in price are exactly offset by differences in retail service. The practice also means that the BLS approach to cross-outlet price linking allows factors other than observed market price to affect the index. The opposite approach would assign all of any change in the price of an item from one outlet to another to real price change.
If changes in consumer outlet choices exhibited no clear trends or if price differences simply reflected the fact that discount stores are achieving lower costs by cutting the quality of services, there would be little reason for concern. However, the last few decades have produced clear patterns of change in shopping
choices. It is well documented that general shifts in shopping habits—specifically, the move from higher-priced full-service outlets to discounters—have occurred. For instance low-price and expanded-format food stores grew from a 31 percent to a 50 percent market share between 1979 and 1988 (Reinsdorf, 1993:228), and discount outlets have permeated well beyond food sectors, into electronics, computer equipment, home improvement, and others. Under the prevailing methodology, a large shift in the types of outlets that consumers choose to patronize does not directly affect the index trend, since all price change arising from rotation is linked out. This potential bias affects major CPI item categories, including food and beverages and apparel, although other categories—most notably housing—are not affected by outlet substitution.
The fact that the market share of low-price discounters has been steadily growing implies that, even after quality adjustment, prices at those stores are lower than elsewhere. If through economies of scale and other means, the large-volume retailers have been able to provide lower prices to a growing number of consumers for the same quality-adjusted goods, the current procedures bias the CPI upward (as a cost-of-living indicator). At the same time, a minority of consumers who would have preferred to continue shopping at traditional stores found them driven out of business by the new outlets; those consumers experience an increase in their living costs.
New outlet bias is lessened to the extent that low pricing at new outlets forces established ones to follow suit while they are still in the CPI sample. Depending on the timing of store pricing responses and outlet rotation, the CPI may capture such a price decline. If the now relatively less frequented outlets that BLS used to sample began lowering prices before they were rotated out, the price pressures created by the new outlets would be captured.20 In fact, within the economic model of perfect competition, lower prices would, in equilibrium, be balanced by poorer quality. If equilibrium were always maintained, there would be no potential for this type of bias. However, it appears that the shift to new outlet types has been an ongoing process that is still continuing. As new outlets open, consumers in the area gradually change their shopping behavior and take advantage of the lower quality-adjusted prices.
The recent emergence of e-commerce (the business-to-consumer component) has the potential to create another disequilibrium situation. Expenditure and sales data indicate that consumers are purchasing a small but rapidly increasing share of goods and services through Internet retailers. BLS is planning to rotate these outlets into its sample more or less according to standard protocol. Current CPI procedures for determining where consumers shop should capture increased
patronage of Internet outlets. Even so, emergence of a new mode of shopping highlights time use, item acquisition, and other considerations that must be thought through in order to make pricing of purchases consistent across outlet types.
To illustrate one dimension of the story—item acquisition—consider the example in which, in period one, a compact disk costs $16 at the local record store. In period two, after outlet rotation, a Tower Records superstore that sells the same compact disk for $15 replaces the local store in the sample. Under current procedures, the drop in price does not figure into the CPI; BLS chalks up this price change to a type of quality difference, implicitly assuming that consumers must bear other costs to obtain the item at the lower price. In this case the new costs may derive from things like longer travel time and increased fuel use to get to the superstore or from sales clerks who cannot provide informed answers about music. Since linking essentially equates the old outlet price with the new, some costs beyond the price paid at the register are implicitly part of the adjustment. BLS assumptions about cross-outlet price variation therefore rely to some degree on the idea that an index should reflect full consumer costs, rather than simple transaction price.
For Internet (or catalog) purchases, BLS includes shipping costs in the price. Extending the example from above, say the compact disk in question costs $14 at Tunes.com but requires an additional $2 shipping fee. The e-purchase will be recorded as $16 for index calculation purposes. If a second outlet rotation pushes Tower out of the sample and brings Tunes.com in, the change in price from $15 back to $16 does not affect the index—it is linked out. The higher Internet purchase price is attributed to a difference in quality between the Tower and Tunes.com outlets and contains no pure price component. As discussed above, it may be questionable to assume that the full price difference is equal to the difference in the value of outlet service perceived by consumers, but the method is consistently applied.
On the other hand, the validity of any price level comparison would be questionable (since this comparison is avoided, this does not initially create any problems). From consumers’ perspectives, the cost of procuring the CD at Tunes.com (shipping) may be viewed as a substitute for time spent (an opportunity cost) and gas used (an explicit cost) traveling to the store. Thus, there may be an element of acquisition cost that is included in the e-purchase that is not in the bricks-and-mortar purchase; the recorded $15 price at Tower does not include the cost to the consumer of driving to the store to buy the CD, while the $16 e-purchase price includes delivery.21
In this example, the value of time, which varies from one consumer to another and which must enter a full-blown COLI concept, becomes a relevant issue since choices relating to it can affect the index. Consider, for instance, that different consumers pay different shipping rates for the same product from the same outlet. Some customers choose to pay an extra $3 to have one CD delivered overnight; others pay only 10 cents for shipping because they order 20 CDs at the same time and choose 2-day ground delivery. Are these consumers paying different prices for the product, or are they paying for time savings or for additional services that contribute in combination with the primary item to their utility? The customer selecting overnight delivery reveals not that he would rather pay $19 than $16 for a CD, but that there is another element besides eventual ownership that affects his well-being. By extension, if everyone changes from buying CDs at the superstore to buying them from the Internet store (and sample rotation reflects this), the price index should not automatically rise (and it wouldn’t under the current linking procedure). However, problems could arise if there is a shift in available delivery options or consumer choices. Say, for example, that BLS originally prices CDs delivered by 2-day mail. Now assume that the Tunes.com upgrades its delivery service by replacing 2-day delivery with overnight delivery without raising prices. If no quality adjustment is made when CD purchases under the new setup are priced, the index would miss a true price decrease. This would clearly be incorrect from a COLI perspective (and, depending on how the good is defined, probably also from a cost-of-goods index [COGI] perspective). The problem could arise if item acquisition cost is not treated in a consistent manner—that is, if in some cases it is left out and in others it (or part of it) is included.
The big question is whether or not any of these outlet effects are quantitatively important to index performance. At present, it seems unlikely, but conditions are changing. For instance, market boundaries are rapidly expanding, which might minimize (or exacerbate) the price dispersion problem. By reducing information and search costs, the Internet may one day make the law of one price assumption less unrealistic. The Internet allows consumers access to a vast amount of product information that enables them to easily shop for the lowest price; this visibility also puts pressure on retailers to match competitors’ prices. Fuller information and global access mean that indexes for, say, San Francisco and Milwaukee might converge toward the national average. Consumers now have fuller access to product specifications and worldwide price information than ever before. Even consumers who ultimately patronize only brick-and-mortar outlets can save time gathering price and product information. The cost of making informed decisions about purchases has decreased; the chance of purchasing at a noncompetitive price has been reduced. The former effect (cost of information) is not captured in current CPI methods; the latter effect (lower price) probably is.
To the extent that e-commerce forces competition, differential price trends for specific geographic areas may be minimized. As such, the increase in e-
market penetration may eventually reduce the relevance of geographic (or demographic group) indexes. In addition, the relevant region of any transaction is becoming more difficult to define. A consumer in Chicago may buy an item produced overseas, sold by a dotcom operating in San Jose, and shipped from a warehouse in Alabama. Theory suggests that elimination of information and geographic market barriers will force retailers toward operating at uniform profit margins. But some unconventional pricing practices within the Internet retailing sector have also emerged: companies that are surviving on market capitalization may operate with large accounting losses, generally in an effort to maximize not short-run profit but market share, though this phenomenon is rapidly perishing with the widespread failure of dotcom start-ups. However, one should not overstate the case for price convergence, given that the evidence so far is surprisingly weak. Part of the reason may be that e-retailers (like catalog merchants) have the ability to price discriminate on the basis of a consumer’s past purchasing behavior or, perhaps, information obtained about the purchaser from other vendors. Furthermore, major CPI components, such as housing and utility services, will always be affected by local market conditions and institutional factors.
Evidence of Outlet Bias
The research attempting to estimate the extent of outlet substitution bias is thin.22 Reinsdorf’s 1993 study, probably the most cited on the topic, formally outlines the underlying theory and offers empirical evidence that outlet price differentials are at least partly real rather than merely reflective of quality differences. Reinsdorf—whose research served as the basis for the Boskin commission’s outlet substitution bias estimates—compared food and motor fuel prices from outgoing and incoming samples during a 2-year overlap period when samples were being rotated. New sample prices were on average about 1.25 percent lower. Given that sample rotation occurs every 5 years—and given the rather strong assumption that the lower prices were not accompanied by a deterioration in service or other outlet-related quality elements—this implies a 0.25 percent annual bias in the relevant components of the index. Reinsdorf provided a second estimate by tracking changes of CPI components against unlinked average price-paid data (also published by BLS). For foods the average price indexes rose 2 percent more slowly than did linked CPI subindexes, and for unleaded gasoline 0.9 percent slower. Since quality change is not controlled for, Reinsdorf asserted that these estimates should be thought of as an upper bound of sorts for the outlet substitution bias.
Lebow et al. (1994) adjusted the Reinsdorf estimate to reflect that only a subset of CPI goods are affected by outlet substitution. The authors determined
that weights associated with the relevant set of categories account for about 40 percent of the CPI, which yields an overall index bias of 0.1 percentage points (0.4 × 0.25) per year. The Boskin commission adopted this estimate in its categorical reporting of CPI outlet bias. Given the absence of alternative evidence, it is hard to fault this choice, but the quantitative effect of outlet substitution remains unclear. Fixler (1993) notes that comparisons between the movement of average prices (as used by Reinsdorf) and the CPI strata counterpart do not provide direct evidence of the outlet effect because “differences in index formula, in treatment of product quality change, and in coverage of average prices and CPI strata indexes” may also play a part in the divergence of the two series (p. 8).
MacDonald and Nelson (1991) also produced a rough estimate of the bias created by the market shift to discount stores by combining information on prices across outlet types and market shares. At the time of their study, data published by the trade journal Progressive Grocer indicated that prices at warehouse food stores were 13.4 percent lower that at traditional outlets. The lower prices, along with a 0.7 percent annual growth of warehouse store market share, imply a non-quality-adjusted 0.1 percent per year index bias. Any quality adjustments in favor of traditional outlets would reduce the bias. Also, bias estimates implied from market share information may be overstated if consumers who preferred traditional outlets are forced, because of the outlets’ extinction, to patronize the less desirable superstores. A full estimate of outlet bias would have to consider the increased quality-adjusted price that traditionalists must now pay.
Estimating the Real Component of Price Differences Across Outlets
To accurately remove outlet substitution bias from a COLI, an index producer must abandon the assumption embedded in either extreme position—that any observed price difference of an identical item at two outlets (1) is wholly attributed to outlet-related quality differences valued by consumers or (2) contains no quality component and therefore reflects pure price variation. In order to escape these assumptions, methods would have to be developed that isolate and quantify the value to consumers of the service, time, and other quality dimensions that differ by outlet type so that the pure price component could be identified. That is, differences in observed price changes associated with outlet-rotated items must be broken down into price and quality components, as is done for items whose embodied characteristics change.23
This prescription certainly suggests a role for hedonic analysis, in this case based on outlet (as opposed to product) characteristics. However, work in this direction is embryonic and applicable results appear to be a long way off. This type of hedonics may be even more complex than in the standard quality adjustment context since such an analysis would need to “allow for the existence of temporary market disequilibria and a distribution of preferences across consumers” (Reinsdorf, 1993:250).24 And the problem of identifying outlet characteristics that are tied to consumer valuations of service is certainly conceptually no simpler than in the parallel item quality case.25
Because any attempt to control for price changes attributable to item and outlet characteristics would require collecting detailed data, scanner technology may offer some hope for advancing research on outlet substitution bias. If outlet identifiers could be incorporated into sales data generated by stores, scanner data could reduce the time needed to track which outlet types are experiencing growing or shrinking market shares. Electronic data sources may also facilitate systematic unit-price calculation to account for different packaging that characterizes small and large shops and for varying promotional sale and coupon use patterns. Scanner data may promote more accurate tracking of actual transaction prices, which would be essential to research on price variation across outlets.
Though current evidence suggests outlet bias is significant enough that index producers should be concerned about it, there is a real question as to whether research can generate sensible, reproducible price estimates that neutralize quality in across-outlet product comparisons. Given the complicated nature of any such estimation, the Conference Board (1999:23) suggested, as an interim solution, splitting the difference between the two extremes—that either all or none of observed price differences are due to quality variation—on the basis that it is no more arbitrary than the “all or nothing” assumption and that is likely to be closer to the “truth.” The board added that such a practice should only be adopted if solid analysis of specific items suggests that the approximation is reasonable. After all, it is certainly possible that one of the two extreme assumptions is in fact empirically closer to the truth than is the half-and-half solution.
Conclusions and Recommendations
BLS treatment of items that enter the CPI through outlet rotation is conceptually parallel to its treatment of items that replace disappearing ones (except that, for outlet substitution, overlap pricing can nearly always be used to impute price change for the transition month). In the case of new outlets, the continuing (for now) shift of purchases toward large discounters suggests that the price differences are not, on average, proportional to perceived quality variation of shopping experiences.
On the assumption that the COGI concept does not call for fixed “weights” among types of outlets then, under either a COGI or a COLI approach, an explicit decision is needed about how to make item price quotes comparable as new outlets replace old ones in the CPI sample. For the foreseeable future, the BLS will not have the tools to explicitly adjust observed prices to account for changing outlet quality characteristics. Thus, the range of short-term recommendations from which the panel may select is limited. BLS could:
continue the current treatment of outlet replacement—in which case all of any price difference for a specific item from one outlet to another is assumed to be equal to the difference in the quality of the shopping experience;
treat the price difference as a “true” price change—in which case zero net quality difference is assumed to exist between outlet types—though this is a poor option because prices at the newly selected outlet were, in most cases, lower before it was rotated into the sample, so even if a (quality-adjusted) price decrease took place, the timing of its inclusion into the index would be wrong; or
split the difference—as recommended by the Conference Board, but there is little reason to believe that splitting the difference is more accurate than the results of the current practice (and such a recommendation might be faulted for creating the precedent of solving a difficult problem without quantitative evidence).
Conclusion 5-2: Given the available options and given that current techniques cannot consistently and accurately separate quality and price effects associated with the value of retail service, BLS has little choice but to continue its current practice.
However, in principle, when outlet rotation results in a change in the observed price of an identical product, an attempt should be made to decompose the difference into quality (or convenience) and pure price components, instead of attributing it, in its entirety, only to the former.
Recommendation 5-2: With longer-term modifications in mind, the panel recommends pursuing research into price variation across outlets with differing characteristics.
As with product-based hedonic techniques, substantial methodological and data advances would be needed before any such changes to the CPI could be put into practice. Undertaking this approach would require BLS to intensify research on consumer search costs, time use, service valuation, hedonic methods applied to outlet characteristics, and, more generally, how to adapt sampling methods to facilitate more extensive quality adjustment across item rotations. Since it seems unlikely that efforts in these areas could have a large effect on the CPI and given their conceptual difficulty, the panel would assign a low priority to this research.