Quantifying Quality Growth

Transcript

1 Quantifying Quality Growth By B ILS AND P ETER J. K LENOW ARK M * Using U.S. Consumer Expenditure Surveys, we estimate “quality Engel curves” for 66 durable goods based on the extent richer households pay more for each good. The same data show that the average price paid rises faster from 1980 to 1996 for goods with steeper quality Engel curves, as if households are ascending these curves. BLS prices likewise increase more quickly for goods with steeper quality Engel curves, suggesting the BLS does not fully net out the impact of quality upgrading. We estimate that annual quality growth averages 3.7 percent ( D12, O40, for our goods, with 2.2 percent showing up as higher inflation. JEL E31) sured quality growth for 66 durable consumer As people get richer they consume not only more goods but better goods. Quantifying goods that constitute over 80 percent of U.S. such quality growth is difficult. Because of spending on consumer durables. Our instru- ment is based on predicting which of these 66 exacting data requirements, the hedonic tech- niques pioneered by Irma Adelman and Zvi goods will display relatively rapid quality growth, then contrasting how unit prices ver- Griliches (1961) and Griliches (1961) have been applied to only a limited number of sus government-measured prices respond to goods (e.g., cars, houses, computers). Mat- these differences in quality growth. Inflation in a good’s unit price reflects growth in the thew D. Shapiro and David W. Wilcox (1996 average quality of the good as well as its true p. 124) describe the measurement of quality change as necessitating “house-to-house com- rate of price inflation. Ideally, the U.S. Bu- reau of Labor Statistics (BLS) fully controls bat,” that is, detailed good-by-good studies. for quality changes, producing measures of The Boskin Commission Report (Michael J. inflation equal to the true rates of inflation. Boskin et al., 1996) cites only a handful of But suppose BLS procedures do not fully studies in arriving at its estimate that unmea- sured quality change biases U.S. Consumer control for quality changes, with part of quality- Price Index (CPI) inflation upward by 0.6 driven price increases inadvertently recorded 1 percent per year. as price inflation. Then BLS inflation rates, We introduce an instrumental variables like unit price inflation rates, will predictably respond to quality increases. In turn, the ex- (IV) approach to estimate the rate of unmea- tent of quality growth that escapes BLS mea- surement can be identified by comparing the * Bils: Department of Economics, Harkness Hall, Uni- magnitude of responses in BLS and unit versity of Rochester, Rochester, NY 14627; Klenow: Re- prices to predictable differences in quality search Department, Federal Reserve Bank of Minneapolis, growth. 90 Hennepin Avenue, Minneapolis, MN 55480. For helpful comments we are grateful to Per Krusell, William Nord- To predict those consumer durables that haus, an anonymous referee, and especially, Matthew Shapiro. will display more rapid quality growth we 1 Including studies on new goods as well as higher exploit “quality Engel curves” that we esti- quality goods, the Boskin Commission cites William C. mate from pooled cross sections of household Randolph (1988) on housing, Robert J. Gordon on durable data (1980 to 1996 U.S. Consumer Expendi- goods (1990), Manuel Trajtenberg (1990) on medical im- aging devices, Griliches and Iain Cockburn (1994) on pre- ture Surveys). Whereas a traditional Engel scription drugs, Steven Berry et al. (1996) on new cars, curve traces out total expenditures on a good David M. Cutler et al. (1996) on heart attack treatment, against permanent income or wealth (which Jerry A. Hausman on breakfast cereal (1997) and cell we proxy with overall consumption), a quality phones (1999), and William D. Nordhaus (1997) on Engel curve traces out the of a good unit price lighting. 1006

2 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1007 2 Our premise is against overall consumption. dices, the quality upgrading that we find re- that, across households at a point in time, flected in unit price changes need not show up those paying higher prices are buying higher- in BLS price changes at all. We find, however, quality goods (perhaps bundled with more that goods with steeper quality Engel curves do display faster rising BLS prices. We estimate retail services). Not surprisingly, richer that, over 1980 –1996, the BLS deflators ad- households do tend to buy more expensive goods, so the estimated slopes are all positive justed for only about 40 percent of the predicted differences in quality growth across goods, with and significant. Averaging across the goods, the quality portion comprises 56 percent of the remaining 60 percent showing up as higher BLS inflation. The BLS netted off a little under the overall Engel curve, suggesting an impor- 1.5 percent per year for quality growth for our tant role for quality growth in consumption 66 goods from 1980 –1996. If this represents growth. only 40 percent of all quality growth during the Our instrument is based on the relative steep- period, then the BLS understated quality growth ness of the quality Engel curves across the 66 goods. For instance, we see that richer house- and overstated inflation by 2.2 percent per year holds buy much more expensive automobiles for our 66 goods. than poorer households do, whereas richer We can briefly summarize our strategy as follows. The data we use have three dimensions households spend only modestly more than of variation: goods, households, and time. For poorer households in purchasing a vacuum each good, we identify its quality Engel curve cleaner. Thus, as households on average be- by regressing the unit price paid by the house- come richer, we predict faster quality growth for automobiles than for vacuums. Assuming hold on the household’s spending on nondura- ble goods. We then identify the fraction of goods with steeper quality Engel curves do not quality upgrading missed by the BLS by re- display systematically faster or slower true in- flation over time, a good’s gressing, for the sample of 66 goods, the time- quality En- relative gel curve provides a valid instrument for average of BLS inflation on the time-average of quality-driven growth in unit prices. unit price inflation, instrumenting for the latter with each good’s quality Engel curve slope. We find that our estimated quality Engel The rest of the paper proceeds as follows. In curve slopes are highly correlated with unit Section I we lay out a simple model in which price changes for the 66 goods (correlation co- efficient of 0.51). That is, those goods with rising household purchasing power generates steeper quality Engel curves display faster ris- rising demand for quality. This model features cross-sectional quality Engel curves specific to ing average unit prices over 1980 to 1996. This is precisely what one would expect if house- each good that provide an instrument for our IV approach to estimating quality growth. In Sec- holds are climbing up their quality Engel curves tion II we present the time-series behavior of over time. We estimate that quality upgrading unit price and BLS price inflation rates for our occurs at the rate of about 3.7 percent per year on average for the 66 goods. This quality 66 goods. In Section III we estimate quality growth can take several forms. One form is slopes for the 66 goods using household data. In Section IV we exploit the quality slopes esti- rising market share of existing, above-average- quality goods. Another is the replacement of mated off of cross-sectional data to predict the existing goods in the market with higher-quality rate of quality upgrading over time, and test the versions. As we discuss below, our methodol- extent to which BLS prices (improperly) rise with quality upgrading. Section V concludes. ogy can in principle capture both types of qual- ity upgrading. I. A Model for Estimating Quality Engel Curves Because the BLS makes explicit adjustments and Predicting Growth in Quality for quality change in constructing its price in- The typical model of quality improvements 2 The overall Engel curve is the product of the quality [see, e.g., Philippe Aghion and Peter Howitt quantity Engel curve, where the latter Engel curve and a (1992)] focuses on firm incentives to design traces out the number of units bought against overall con- sumption. higher-quality goods. The preference side of the

3 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1008 model is usually kept simple, with consumers - ̃ n veys of U.S. households. captures house iht i hold h ’s taste for good at time preferring higher quality but substituting with in- t . The parameters s finite elasticity among different qualities. We will govern the curvature of utility for the and s i s 0 and . present evidence that, in contrast, different levels goods, and we assume s 0 @ i . . i We abstract from uncertainty, allowing for a of quality are imperfect substitutes in the eyes of constant growth rate of real expenditures. We consumers. Richer households typically buy more has a deterministic life of expensive, higher-quality versions of goods. In t i assume good i periods. Therefore, a household owns good if i this section we lay out a simple model that has this it purchased the good in this or one of the feature. We derive quality Engel curves that relate t preceding ( the quality of good purchased (measured by price 2 1) periods. We do not treat t i i as a choice dimension of quality. We assume paid) to a consumer’s wealth and consumption. In t consumers keep the good for the full turn, the relative slopes of the quality Engel curves periods. i predict which goods should exhibit faster rates of Thus consumers do not trade in used goods, which we think is realistic for most of the goods quality improvement over time. we examine. This requires that the desired growth in quality over the life t A. Household Quality Choices of a good is not i so fast that consumers would choose to discard a working durable to upgrade its quality. At time 0, household h maximizes lifetime The household budget constraint is utility given by N ` t , y 5 x V c U 5 1 b (1) u , O O h ht 0 ht iht ht iht 1 5 i t 5 0 is the discount factor. u b where where is utility ht derived during period t : z q . 5 (2) x it iht iht s 1 2 1/ 1 c 2 ht 5 u ht In (1) the price of nondurable consumption is 1 s 1/ 2 y is household expendi- normalized to one and s 1 1/ 2 ture, which equals income minus the change in i 1 @ # q 2 N iht assets. V - h is 1 if household purchases dura if q . 0 ̃ n iht iht iht 1 s 1/ 2 1 O i ble i in period t , and 0 otherwise. x unit is the H iht 1 i 5 0if 0. q 5 iht for good h price in period i paid by household . As shown in (2), the unit price is the product t of good of the common i quality-adjusted price Each household chooses q , the quality of iht ( facing all households at time t z ) and the , for good different durable, indivisible i N it of good at time h bought by household i quality goods. A household may choose not to own q t ( ). This captures the idea that, for a given t , in which case at time i q durable good 5 iht iht 3 (say televisions), the house- i type of product 0. Household h also buys an effective amount hold faces a menu of quality–price combina- c (quality times quantity) of the divisible, ht tions from which to choose. The menu slopes composite nondurable good. We separate out upward, so that higher-quality versions are indivisible goods because these are the ones for 4 We assume that the relative more expensive. which “unit prices” (the price paid for a unit of the good, such as for a single refrigerator) are observable in the Consumer Expenditure Sur- 4 - In (2) we define quality in price terms, so that a dou bling of quality doubles price. Our results are robust to 3 assuming a more general elasticity f Subtracting 1 inside the brackets means utility from the of price with respect i f i to quality (i.e., x q good is positive only if q . 5 1; that is, it is not worth ). What is important for the z iht i i consumer’s problem is the extent of diminishing returns to buying the good unless one buys a sufficiently high-quality version. This contributes to some households not owning spending on quality. These diminishing returns can reflect certain goods at all. This functional form also allows utility either diminishing utility flow from quality because s ,` , i to be positive even when or a rising price of quality from s f 1, given q . 1. , . 1. i iht i

4 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1009 price of differing qualities of a good are deter- a household’s demand for quality (expressed in price units) rises as its consumption of nondu- mined by relative production costs, given competitive pricing. Moreover, this rate of trans- i rables rises. Good ’s quality slope is steep if formation between lower- and higher-quality ver- there is little curvature in preferences with re- q spect to sions is unaffected by relative or total quantities is high). The quality slope s (i.e., if i i produced. [Sherwin Rosen (1974) considers is important not only for how quality responds somewhat more general assumptions.] to nondurable consumption, but also for how Facing the quality–price menu, each house- quality responds to shifts in the quality-adjusted i . Suppose the cost of producing price of good hold chooses whether to buy a good and, if so, what quality level to buy. We focus on the latter good z increases 1 percent, raising i - by 1 per i q cent. If there is no response in the level of decision, treating quality as a continuous i being rises by 1 i quality bought, the unit price of choice variable. Conditional on good i percent. But, this increase in z purchased, the household equates the ratio of will induce the i - q marginal utilities of su purchased to fall by i quality of good (derived over the subse i i percent. quent t c to the ratio of their periods) and i prices: B. Predicting Growth in Quality t i 1 2 b s 1/ 2 i n ̃ q S D iht iht We draw a distinction between how quality 1 2 b upgrading affects inflation in unit prices versus z (3) . 5 it s 2 1/ c ht BLS prices. The growth rate of unit prices re- flects the sum of quality growth and “true in- flation” (the growth in prices holding average Rearranging and taking natural logs yields quality constant): c ln n ln 1 , z ln su 2 q u 5 (4) ln iht iht it ht i i 5 x q 1 D D z , (5) D i i i where D where x denotes the growth rate (i.e., log first i difference) of x . Expression (5) derives from su t i i ~ 1 2 b n s ̃ ! i iht averaging log first differences of (2) across buy- n u 5 and 5 . D S i iht s b 2 1 ing households. The overbars denote time aver- ages, which we use to emphasize that the empirical implementation will involve time- Expression (4) shows that, conditional on buy- i ing good averages of inflation rates (specifically, over , a household will choose a higher- 1980 –1996). As shown in (5), the only varia- quality version the richer is the household (the higher is c tion remaining is across goods . i ), the lower is the quality-adjusted ht z price of the good (the lower is In contrast to unit price inflation rates, BLS ), and the it greater is the household taste for the good (the inflation rates aim to measure price changes higher is n We denote the BLS holding quality constant. ). iht D p as i inflation rate for good From (4), the elasticity of demand for quality . If the BLS i with respect to c is u measure is unaffected by changes in quality, i for good . We call this i then it equals D z the slope of the “quality Engel curve” for good . If, instead, the BLS is able to i 2 m ) of quality , or “quality slope” for short. It maps out how net out only a fraction (1 i growth, then p D is given by i relative We have also considered the possibility that the D (6) D 1 m D q p . z 5 i i i price of quality for good i rises or declines over time through changes in the parameter f . Changes in f will be i i reflected in a changing slope of the quality Engel curve If the BLS deflator perfectly measures price per discussed below. We find, however, that for most of our 66 unit of quality, then is zero. If the BLS m goods we cannot reject constancy of the quality Engel curve from 1980 to 1996. (See Section IV.) understates quality improvements and overstates

5 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1010 inflation, then is positive. As stressed by Jack m slopes for 66 consumer durables using cross sections of the Consumer Expenditure Sur- E. Triplett (1997), however, the BLS might vey. We find important differences across overstate quality improvements and understate is negative. inflation, in which case m goods in their estimated quality slopes. Fur- thermore, these differences turn out to be Combining (5) and (6) yields the following relation between BLS and unit price inflation: excellent predictors of which goods display faster unit price inflation. The correlation be- tween a good’s quality slope and its average ! D z . (7) 1 p D ~ m 5 1 x 2 m D i i i rate of unit price inflation is 0.51. The relevance of the quality slope as an in- m —the fraction of x D Our strategy is to estimate strument for does not guarantee its validity, i that is, its orthogonality to the error term (1 2 quality growth that goes unmeasured— by re- m ) D z gressing BLS inflation on unit price inflation, in (7). Our identifying assumption is that i z m differences in the estimated quality slopes ) as in (7), treating (1 D 2 as an error i term. Now, as (5) shows, unit price inflation is u across goods ( values) are uncorrelated i z D clearly correlated with true inflation with quality-adjusted relative price shifts across .The i key is to instrument for unit price inflation z D goods ( values): i with variables that predict a good’s rate of quality upgrading but are arguably orthogonal ~ u ! , D z (10) Cov 5 0 across i . i i to its true inflation rate. We exploit differ- ences across goods in the slopes of their qual- If (10) holds, then u is a valid instrument for i 5 ity Engel curves (their u values) to construct D x We provide evidence in Section IV in (7). i i these instruments. to support this identifying assumption. For ex- Taking first differences of (4) and averaging ample, we show that factor prices did not rise across households and time, the growth rate of faster, nor did total factor productivity (TFP) is given by i quality demanded for good grow slower, in the industries producing goods with steeper quality slopes. The conjectured relationship between the unit D D . n D (8) 1 q z 5 u D su 2 c i i i i i price x , the BLS price , and nondurable p i i c is depicted in Figure 1 for two consumption This says that goods with steeper quality slopes goods (vacuums and cars). For each good, the (higher values of u ) should exhibit faster i unit price, quality-adjusted price, and BLS price growth in quality in response to economywide income and consumption growth ( are normalized to equal each other in the base . 0). c D Quality should also rise faster for goods with period ( x ). Growth from period 0 p 5 z 5 0 0 0 declining relative prices ( D to period 1 in nondurable consumption gener- z 0), particularly , i if the good has a steep quality slope. ates an increase in quality and unit price for u i good Substituting (8) into (5), unit price inflation equal to . The figure is drawn such c D i u equals that is larger for cars than for vacuums; cars i exhibit the relatively steeper quality slope. For this reason, the increase in x x from is x to 1 0 i 1 1 2 su D ! D z c 1 D n (9) . 5 D x u ~ much larger for cars than that for vacuums. If i i i i i the BLS price reflects only quality-adjusted prices, then the growth in p , p , from p to 0 1 i The first term in (9) says that goods with steeper quality slopes display faster average 5 More formally, the condition is growth in unit prices in response to economy- wide consumption growth, reflecting their N N 2 ! z ln z ln u ~ • 1 2 it k it i 5 i 1 faster growth in quality. This means that dif- lim u 2 O l G F N Nk ferences across goods in the quality slopes u ` 3 Nk i l 5 1 should be a relevant instrument for differ- N ences in unit price inflation rates D x across i 1 z 0. 3 z ~ ln ! ln 2 5 goods. Below we estimate separate quality O lt lt 2 k G F Nk 5 l 1

6 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH 1011 VOL. 91 NO. 4 N u of z D z 5 (1/ N ) • across goods and D 5 1 i i i is the average true inflation rate across goods. This expression suggests the interaction term ( u as a second relevant instrument for z 2 u ) D i i the growth rate of unit prices. It captures the feature that quality will respond most dramatically z to a change in for a good with an especially steep i quality slope. Validity of ( u D z 2 u ) as an i i instrument requires an assumption that parallels (10), but with the quality slopes uncorrelated 2 with ( D z D . rather than with ) z i i Construction of the instrument ( u D ) u 2 z i i is complicated by the fact that D z is not directly i observable. We observe the BLS inflation rates, C F NGEL 1. C LIMBING U P Q UALITY E URVES IGURE but these are equal to the true inflation rates RICES P DJUSTED -A UALITY Q OLDING (H ) ONSTANT C 0. Rearranging (7), true inflation 5 m only if qz ; z 5 5 p 5 Notes: x unit price 5 quality-adjusted price; can be related to BLS inflation and unit price m BLS price 5 q . z inflation by 1 should not be greater for cars than that for D m 2 p D ~ x 5 z D (12) . ! i i i 1 2 m vacuums. Figure 1 depicts no changes in quality- D z adjusted prices ( 0) for both cars and 5 i This construction requires a value for m vacuums, so the BLS prices should not change , the parameter of interest. Therefore, its use in form- at all. But to the extent that m is greater than zero, faster quality growth in x ing another instrument entails nonlinear estima- for cars than for i . We return to these issues in Section vacuums will be mirrored in faster growth in tion of p m . i IV. As drawn, about two-thirds of the faster , we can estimate m Given an estimate for growth in the quality and unit price of cars relative to vacuums shows up as faster BLS quality growth and unmeasured quality growth for our set of consumer durables. If the BLS inflation for cars. This would identify a value 2 . ⁄ for succeeds in fully netting out the impact of qual- of m 3 Of course, quality-adjusted prices do change ity change, then quality growth is simply the over time, and at different rates for different growth rate in unit prices for good minus its i BLS rate of price increase. When goods. We do not rule out such shifts in our 0, how- m . IV estimation of m equals . We assume only that the ever, quality growth for good i shifts are orthogonal to the quality slopes iden- tified off of cross sections of households, as 2 x D p D i i expressed by condition (10). Moreover, we ac- (13) 5 . D q i m 2 1 tually utilize changes in quality-adjusted prices to construct another instrument for quality unmeasured The extent of growth. To see this, first rewrite (9), ignoring quality growth for is similarly given by good i constant terms, as D (11) ! 5 u D ~ x c 2 z s D i i p D 2 x D ~ m ! i i 2 ! 5 . x D ~ D p D q (14) 2 i i i 1 2 m D 2 z ! z D s ~ u !~ 2 u 2 i i We highlight two limitations of our approach n , su ! D z 1 D ~ 1 1 2 i i here. First, we are making the strong assump- tion that —the extent of quality growth that m N u N where seeps into BLS inflation rates—is the same ) (1/ • 5 u is the average value 1 i i 5

7 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1012 ERCENT B UYING I NDIVIDUAL C ONSUMER G OODS IN A T YPICAL Y EAR T 1—P ABLE (3) (1) (2) Percent buying 2 1 Number buying Good (of those buying) Percent buying 7.4 4,835 Carpeting 19.6 Curtains and drapes 20.1 9,251 14.2 10.5 9.1 5,911 Mattress and springs 10.2 6,649 Bedroom furniture 17.2 Sofas 5,347 8.2 8.4 Living room furniture 13.4 30.5 8,731 12.5 7.9 5,131 Kitchen/dining room furniture 4,915 7.5 35.8 Baby furniture and equipment 5,731 Outdoor furniture 8.8 14.2 12.9 4,365 Refrigerators and freezers 6.7 10.3 4.9 3,205 Clothes washers Clothes dryers 3.4 11.4 2,235 Stoves and ovens 2,563 3.9 15.1 5.5 Microwave ovens 3,567 5.2 Window air conditioners 2.2 5.1 1,435 15.9 10,346 Televisions 11.0 15.2 8,224 Radios 12.6 Stereos 4,953 7.6 12.7 5,757 15.6 8.8 Rugs 8.1 5,256 Window coverings 14.2 Clocks 5,218 8.0 11.3 Lamps and lights 8,695 13.3 18.9 a Telephones 9,379 14.4 18.6 12.4 8,112 Lawn and garden equipment 20.0 Power tools 6,247 9.6 25.6 8.7 Vacuums 5,045 7.7 4.4 1.8 Sewing machines 1,202 20,270 31.1 33.6 Small kitchen appliances 12.0 Heaters 6,530 10.0 33.8 1,088 Hard flooring 1.7 Office furniture 3.5 13.8 2,311 15.8 10,298 Hand tools 34.5 Men’s suits 23.1 8,663 13.3 Men’s coats and sportscoats 18,837 28.9 35.9 9,592 28.9 14.7 Men’s and boys’ sleepwear 28.2 18,378 Men’s and boys’ sweaters 40.3 Men’s pants 34,812 53.4 55.8 14.0 Boys’ coats, suits, and sportscoats 9,124 44.0 Women’s and girls’ coats 41.5 47.1 27,068 57.9 Women’s and girls’ dresses 34,502 52.9 Women’s sweaters and vests 48.9 26,358 40.4 65.3 38,565 Women’s skirts and pants 59.2 48.4 Women’s and girls’ sportswear 21,695 33.3 34.5 41.0 Women’s sleepwear 22,475 29.9 17.4 11,373 Women’s suits 47.5 47.1 Men’s footwear 30,682 76.6 20,525 31.5 Boys’ and girls’ footwear 63.3 41,274 Women’s footwear 62.8 17,489 Watches 26.1 26.8 55.5 39.0 25,439 Jewelry 19.4 Luggage 6,614 10.1 13.5 13,483 Cars 20.7 7.2 Trucks 4,489 6.9 25,597 39.3 34.7 Tires 32.6 Eyeglasses and contacts 18,901 29.0 47.3 26.1 Sports and exercise equipment 16,989 Bicycles 5,401 8.3 19.5

8 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1013 T —Continued. 1 ABLE (3) (2) (1) Percent buying 2 1 Good Number buying Percent buying (of those buying) 28.9 5.0 3,237 Camping equipment 45.4 6,903 10.6 Fishing and hunting equipment Winter/water sports equipment 34.5 10.0 6,523 1.9 10.5 1,263 Playground equipment 7.4 4,814 36.5 Musical instruments Photographic equipment 6,665 10.2 17.3 Personal care appliances 10,389 15.9 25.3 b Calculators 11.3 4,625 7.1 a Typewriters 2.5 5.6 1,610 11,321 26.6 Mean 17.4 Median 6,784 10.4 20.1 Standard deviation 9,911 17.1 15.2 76.6 Maximum 63.3 41,274 1,088 4.4 Minimum 1.7 Notes: Sample: Cross sections of households in the 1980 –1996 U.S. Consumer Expenditure Surveys. Observations: 65,189 household-years. Fraction buying: percentage of households buying 1 or more of the good in a 12-month span. Fraction buying 2 1 (of those buying): percentage of buying households who buy more than 1 in a 12-month span. a 1983–1996. The 1980, 1981, and 1982 Consumer Expenditure Surveys did not include this item. b 1982–1996. The 1980 and 1981 Consumer Expenditure Surveys did not include this item. II. Comparing Data on Unit Price Inflation and across goods. Our cross-good estimation meth- BLS Price Inflation odology does not afford good-by-good esti- 6 mates of quality growth or quality bias. A. Consumer Expenditure Data Second, the overbars indicate time-averages, meaning we do not produce period-by-period estimates of quality growth and quality bias. We We construct measures of unit price inflation for each of 66 consumer durables based on avoid higher-frequency estimates for a couple of reasons. As we describe in the following household spending reported in the 1980 to section, the number of annual unit price obser- 1996 Consumer Expenditure Surveys (CEX) 7 As discussed in the conducted by the BLS. vations per good in the CEX renders the annual growth rates sufficiently noisy that we see it as next section, we also use cross sections of the CEX as our data for estimating quality Engel preferable to time-average the growth rates over curves for each of the goods. the entire 1980 –1996 sample. We are also con- cerned that a demand shift toward goods with The CEX has a rotating sample of about steep quality slopes may lead to a short-run 5,000 households. Each household is main- tained in the sample for a year, encompassing relative increase in factor prices for those prod- four quarterly surveys. The CEX asks respon- ucts. If so, at cyclical and higher frequencies, dents how much they spent over the previous this would go against the identifying assump- tion that differences in the quality slopes are quarter on a wide array of goods and services. Expenditures are typically assigned to a partic- uncorrelated with quality-adjusted relative price shifts. ular month in the quarter. If an expenditure can 6 7 The BLS conducts two separate surveys of consumer We can allow for limited differences in the value of m expenditures, an interview survey and a diary survey. Our across goods. For instance, in Section V we explore the data are based on the interview surveys. We obtained the is m possibility that measurement is more accurate (i.e., 1980 –1994 data from the University of California, Berkeley closer to zero) for goods with greater expenditure shares or (2000) and the 1995–1996 data from the U.S. Bureau of goods for which the BLS sometimes employs direct quality Labor Statistics (1998, 1999). adjustments.

9 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1014 be associated with a particular unit purchase, then fraction of those purchasing a good report more than one purchase in the 12-month period. This we can assign a unit price to the purchase of that fraction is highest for boys’ and girls’ footwear. good. From all the goods surveyed by the CEX, we chose 66 goods for which purchases tend to be 8 We were also restricted by the quite distinct. B. Unit Price Inflation requirement that the BLS produce a price deflator for the good for all or much of the 1980 to 1996 We measure increases in unit prices for the 66 goods as follows. Expenditures are grouped period. The goods are listed in Table 1. by year of purchase. We then construct for each These 66 goods constitute 81.3 percent of a household’s spending on durables as reflected good the average price paid across households 9 Across the 66 goods in the December 1997 weights for constructing by year for 1980 to 1996. the CPI. They represent 12.4 percent of the we have 1,469,561 unit price observations. We overall CPI. [We report the CPI weight for each then divide each unit price by the CPI for non- 10 good in column (1) of Table 3, which we dis- To durables in the same year (our numeraire). minimize the impact of outliers in a particular cuss further below.] The first column of Table 1 reports, for the year, we calculate a three-year centered moving average of these prices. Finally, we calculate the pooled 1980 to 1996 cross sections, the number of households purchasing each good. These annual percentage rate of inflation for each good numbers provide the sample sizes for estimating based on comparing this moving average for 11 the quality slopes in Section III. The second 1995 to its value for 1981. The resulting inflation rates appear in the first column presents the fraction of the sample buy- column of Table 2. Weighting by importance in ing. This ranges from a low of 1.7 percent for sewing machines to a high of 63.3 percent for the CPI, average unit prices rose by 0.97 percent women’s footwear. The final column reports what per year (relative to the CPI for nondurables) on average across the 66 durable goods. The most extreme declines were for microwave ovens ( 9.2 percent) and heaters ( 2 2 4.1 percent). The 8 If a respondent purchases more than one of the same most extreme increases were for trucks (3.7 category of good in the same month (e.g., bicycles) the percent), sports and exercise equipment (2.8 survey may report them separately. But it is conceivable that the amounts can be lumped together. If so, then our quality percent), and jewelry (2.8 percent). Engel curve estimates may be biased upward. This does not compromise the validity of our instruments, however, un- C. BLS Inflation less any such bias from lumping purchases happens to be more important for goods that experience faster true inflation. BLS prices are not the same as CEX unit For the years 1994 to 1996 the CEX asks households to prices for a number of reasons. One important state explicitly the number of items purchased for each of the clothing categories, as well as for watches and jewelry. Thus for years 1994 to 1996 we can compare these re- 9 sponses to the quantities we obtain by summing the number Expenditures are weighted by a CEX sampling weight of itemized purchases in each category of goods. For these for each household. For 12 of the 66 goods we actually goods we find a tendency for our base calculations to calculate inflation rates at a slightly finer level of aggrega- tion than that in Table 1. For instance, living room furniture understate somewhat the number of goods purchased, con- sistent with some lumping. Of much more relevance to our is separated into tables versus chairs; men’s and boys’ sleepwear, as well as sweaters, are separated for men’s work, however, the extent of this discrepancy is typically versus boys’; winter sporting goods are separated from only very weakly related to household nondurable con- water sporting goods. We aggregated goods in these 12 sumption (and hence will have little effect on the quality slopes we estimate below). Based on these comparisons for cases to be consistent with BLS categories. We aggregate on the basis of expenditure shares in the CEX. Similarly, in years 1994 to 1996, we rescale the quantities for each of the clothing categories, watches, and jewelry to correct for the Section IV the quality Engel curves for these 12 goods are extent our quantities systematically deviate from the re- estimated including a dummy variable to control for the sponses to the more direct question on number of items finer category of good being purchased (e.g., is the good purchased. We also condition on family total nondurable men’s sleepwear or boys’ sleepwear). 10 consumption, as well as additional controls (e.g., age of We obtained all BLS price deflators from the BLS web site (U.S. Bureau of Labor Statistics, 2000). household head), in rescaling these quantities. These cor- 11 rections also modify the unit prices. Our results are not For two of the goods, calculators and typewriters, data sensitive to these small adjustments. begin in 1982; for telephones data begin in 1983.

10 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1015 reason is that the BLS collects prices on goods percent of the nonapparel inflation rate in those years (see their Tables 5 and 6 pp. 338 – 40). at a finer level of detail than the CEX categories These figures indicate that item substitutions and leaves the weight on each item unchanged coincide with disproportionately rapid BLS in- from period to period. In contrast, average unit prices reflect current (and therefore changing) flation and, perhaps, unmeasured quality 12 improvements. If people switch toward more expen- weights. sive models within a CEX category, then the The item-substitution rate is even higher for average unit price for the category should rise, the consumer durables that we examine than for although the BLS price index for the category the average item in the CPI. Column (2) of need not. The BLS fixed weighting scheme Table 3 contains the monthly item-substitution 13 The means it does not register a price change when rates in 1997 for the 66 goods we study. substitution rate varies from 2.4 percent per consumers switch among items with different, month for calculators and typewriters to 38.3 but themselves unchanged, prices. This is true even if the BLS collects prices on only a single percent for women’s and girls’ dresses, and averages 13.8 percent across the goods when model in a CEX category. each good is weighted by its share of the De- Although the fixed-weight scheme could pre- cember 1997 CPI [the weights are given in vent quality upgrading from contaminating BLS price changes, the protection is not complete column (1)]. In contrast, the monthly substitu- because many models disappear, forcing the tion rate for all items in the CPI was 3.8 percent in 1997. BLS to price different items from one period to Conditional on the need for an item substitu- the next. The items that disappear may be re- tion, the BLS follows one of three procedures. placed with higher-quality goods, and the asso- In roughly one-half of substitutions (see Sha- ciated quality improvements may not be fully piro and Wilcox, 1996 p. 99) the BLS finds a netted out from the BLS inflation rate. Moulton and Karin E. Moses (1997) describe BLS “item replacement item it judges to be “comparable” to the old item, and makes no quality adjust- substitution” procedures in detail. They report ment. Column (3) of Table 3 reports the per- that about 30 percent of BLS items disappear at centage of substitutions judged comparable for least once every year (p. 323). Moreover, in the three years that have been studied, replacement our goods. It is the most common procedure, items contributed disproportionately to the occurring 46 percent of the time for our goods overall CPI inflation rate. Even excluding ap- (weighted by their CPI share). For certain cat- egories the BLS makes a direct quality adjust- parel, in which items tend to get marked down ment, involving either hedonic pricing or the before being replaced by full-priced items, re- manufacturer’s estimate of the cost of produc- placement items represented 2.6, 2.7, and 3.2 ing the new item relative to the displaced item. percent of price quotes in 1983, 1984, and 1995, Column (4) reports that this occurs 22 percent respectively, but accounted for 20, 34, and 31 of the time for our goods. It is most common for trucks, cars, and men’s suits. For the rest of the 12 In 1996 the BLS collected price quotes for goods in substitutions the BLS scales the entry price of around 200 categories, most corresponding to the CEX the replacement item so that the item’s inflation categories. On a monthly basis, they collected about rate matches that of other items in the same 100,000 price quotes across 44 geographical areas. Accord- category for that month. This usually entails ing to Brent R. Moulton (1996), the mean number of price quotes per category area was 13 in May of 1996. There were scaling the entry price down, and therefore net- not 13 distinct models per category, however, because some ting out some of the higher price of the new were the same model at different outlets. The BLS does not good as reflecting superior quality. Column (5) tabulate the number of distinct models for which prices are of Table 3 reports that this procedure was used collected per category. in 32 percent of item substitutions for our A more minor distinction between BLS and unit prices is that the BLS updates the establishments at which it collects goods. Thus, for the majority (78 percent) of the prices only every five years. Thus a shift toward, say, discount outlets would tend to make CEX unit prices rise more slowly than BLS prices. Both Shapiro and Wilcox 13 - (1996) and the Boskin Commission (Boskin et al., 1996) We obtained this item-substitution data from Appen estimate such “outlet bias” to be about 0.1 percent per year. dix VIII in U.S. General Accounting Office (1999).

11 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1016 HANGES IN U NIT VERSUS BLS P RICES T 2—C ABLE (3) (1) (4) (2) Subperiod Column (2)– Subperiod 1980–1996 annual percent annual percent column (3) Annual percent 5 implied BLS change in unit change in change in unit ( Good BLS prices prices prices quality change) 3.86 2 2.17 2.17 Carpeting 1.69 0.35 0.41 0.06 2 0.35 Curtains and drapes 0.29 0.37 2 0.36 2 0.37 2 Mattress and springs Bedroom furniture 1.09 1.09 2 0.17 1.26 0.95 0.31 2 0.64 2 0.64 2 Sofas 2 0.60 0.51 Living room furniture 2 0.09 2 0.09 2 2 0.76 Kitchen/dining room furniture 2 0.58 0.58 1.53 4.10 2 1.60 1.60 Baby furniture and equipment 2.51 0.60 2 0.06 2 Outdoor furniture 0.66 2 0.66 2 Refrigerators and freezers 1.90 0.43 2 1.47 2 1.47 2 2.53 2 2.60 0.07 2 2.53 2 Clothes washers 1.70 2 Clothes dryers 0.26 2 1.70 2 1.96 2 2 1.85 0.77 Stoves and ovens 2 1.39 1.08 7.81 9.22 2 1.55 2 Microwave ovens 2 2 6.26 2 2 2.19 2 Window air conditioners 1.81 0.48 1.32 2 4.68 1.67 2 1.67 2 Televisions 6.35 Radios 4.62 2.27 2 2.35 2 2.35 2 3.13 2 1.70 2 1.43 2 3.13 2 Stereos 1.06 2 Rugs 2 0.68 2 0.01 2 0.38 2 2 0.24 2 0.75 Window coverings 2 1.00 1.00 2 1.43 0.33 Clocks 2 1.10 1.68 1.26 2 0.95 2 0.30 Lamps and lights 2 1.26 2 2 0.75 5.69 Telephones 2 1.03 4.66 1.40 1.94 0.54 0.54 Lawn and garden equipment 2 2 0.29 2 0.40 0.11 Power tools 2 0.29 2 2 2 1.46 2 0.12 Vacuums 1.59 1.59 0.21 2 0.64 Sewing machines 2 2.70 0.43 0.84 1.32 1.35 2 Small kitchen appliances 2 2 2.19 2 2 0.60 4.09 2 Heaters 1.44 2 2.04 2.52 2 2.49 2.49 Hard flooring 0.02 0.17 2 1.42 Office furniture 0.33 2 1.59 2 2 Hand tools 1.01 2 0.12 0.46 0.58 0.58 2 0.32 0.32 Men’s suits 0.26 0.45 2 Men’s coats and sportscoats 0.18 2 0.45 2 0.63 2 2 1.04 0.96 Men’s and boys’ sleepwear 2 0.08 0.08 2 2 0.35 0.77 Men’s and boys’ sweaters 0.14 0.63 1.41 2 0.44 0.44 Men’s pants 0.98 1.81 2.53 Boys’ coats, suits, and sportscoats 0.71 0.71 2 1.65 2 1.77 0.12 Women’s and girls’ coats 1.65 2 2 1.73 2.30 0.58 0.58 Women’s and girls’ dresses 2 2 0.92 Women’s sweaters and vests 2.70 1.78 2 0.17 2 0.36 0.37 2 3.10 2.73 Women’s skirts and pants 2 2 2 2 1.44 0.71 Women’s and girls’ sportswear 0.73 0.73 2 0.70 0.44 1.54 2 Women’s sleepwear 2 2 2.24 0.04 0.74 0.69 2 0.04 2 Women’s suits 2 0.76 Men’s footwear 0.20 0.20 2 0.96 2 1.63 Boys’ and girls’ footwear 0.31 0.31 1.32 1.61 1.71 0.10 0.10 Women’s footwear 2 2 0.68 Watches 0.41 2 0.27 2 1.29 2 1.58 0.30 Jewelry 2.75 1.28 3.06 2 0.02 Luggage 0.19 2.86 2 2.10 1.75 1.75 Cars 0.35 0.16 0.67 Trucks 3.73 0.83 2 0.94 2 3.19 2.25 Tires 2 0.94 0.49 0.20 2 0.69 0.44 Eyeglasses and contacts 2

12 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH 1017 VOL. 91 NO. 4 —Continued. 2 ABLE T (1) (2) (3) (4) Subperiod 1980–1996 Subperiod Column (2)– annual percent 5 Annual percent column (3) ( annual percent change in unit change in implied BLS change in unit quality change) prices prices BLS prices Good 2.36 2.77 Sports and exercise equipment 5.46 3.09 2 2 1.59 2 1.59 2 1.76 0.17 Bicycles 0.47 0.36 Camping equipment 2 0.12 0.59 Fishing and hunting equipment 1.24 1.23 2 0.09 1.32 1.59 2 2.49 4.08 Winter/water sports equipment 2.49 1.54 1.36 2 0.55 Playground equipment 2 2.90 2.12 2 0.00 2.13 2 2.13 2 Musical instruments 2 0.16 2 0.95 2 0.95 2 0.79 Photographic equipment 0.85 2 0.77 2 0.77 2 1.62 Personal care appliances Calculators 2 1.22 0.28 2 4.02 4.30 2 Typewriters 1.06 1.83 2 0.77 2 2.47 Mean 2 0.44 2 0.39 2 0.94 1.33 2 0.41 2 0.36 2 Median 0.68 1.21 1.74 Standard deviation 1.86 1.53 1.57 3.06 Maximum 2.49 3.73 5.69 7.81 2 9.22 2 2 6.35 2 2.90 Minimum Weighted mean 1.46 0.64 2 0.97 0.82 Notes: The “unit price” is the average of all purchases made in each year across households. The unit prices for the 66 goods are based on 1,469,561 price observations. The period is 1982–1996 for calculators, and 1983–1996 for telephones and typewriters. The weighted mean is calculated using the CPI shares in 1997. Subperiods are because the following years were not covered by the BLS price series: 1980 –1981: Stoves and ovens; microwave ovens. 1980 –1982: Window air conditioners; small kitchen appliances; heaters; hand tools; womens’ skirts and pants; womens’ sleepwear; girls’ coats and jackets. 1980 –1983: Rugs; clocks; mens’ and boys’ sweaters; womens’ sweaters and vests; trucks. 1980 –1984: Luggage; sports and exercise equipment; playground equipment. 1980 –1985: Telephones; hunting and fishing equipment; calculators; typewriters. 1980 –1986: Watches; jewelry; eyeglasses and contacts. 1990 –1996: Sewing machines. 1992–1996: Camping equipment. 1993–1996: Microwave ovens. item substitutions for our goods, the BLS made none of it would go unmeasured. (This is a no direct quality adjustment. This underlines the “passive” quality adjustment that rightly occurs as a result of the fixed BLS weights.) Now possibility that many item substitutions could suppose that, because of rising demand for qual- involve unmeasured improvements in quality that should have been (but were not fully) netted ity, the makers improve the quality of each model in a new year. Item substitutions should out of the BLS inflation rate for those goods. An example may be useful to illustrate these then be triggered. For cars the BLS sometimes makes “active” or direct quality adjustments, ideas. Suppose a particular Toyota Camry is included among the items in the CPI, as is a but this is not typical for all goods or even for our set of durable goods. Whether direct adjust- more expensive Lexus. Suppose further that these car models remain unchanged from one ments are made or not, however, the possibility year to the next, but that households become arises that item substitutions are associated with richer so that unit sales of the Lexus rise relative quality upgrading that is not entirely netted out in BLS inflation calculations. to those of the Camry. No item substitutions need occur. The BLS, by putting a fixed weight Buying improved models that hit the market, on each model across the years, will register no as in this car example, may be an important way inflation at all from the quality upgrading. In in which quality growth occurs over time. Our this example there would be quality growth, but quality slopes, although estimated off of cross-

13 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1018 TEM S UBSTITUTIONS AND M ETHODS OF Q UALITY A DJUSTMENT , 1997 T 3—BLS I ABLE (3) (4) (1) (2) (5) Percent 1997 item- Percent of weight in Percent of Percent of substitution rate (in substitutions substitutions substitutions December “linked” Good percent) “direct” “comparable” 1997 CPI 89 0 11.8 0.021 Carpeting 11 Curtains and drapes 6.5 57 0 43 0.052 0 51 Mattress and springs 0.158 6.9 49 0.203 64 0 36 6.3 Other bedroom furniture 50 Sofas 0.225 7.7 50 0 0 67 Living room tables 0.179 4.5 33 Kitchen/dining room furniture 68 32 7.0 0.146 0 38 0.044 Baby furniture and equipment 0 62 8.0 Outdoor furniture 0.025 19.4 62 0 38 0 9 91 9.6 0.084 Refrigerators and freezers 0 0 Clothes washers 0.057 8.5 100 0 0.035 Clothes dryers 5.4 94 6 0 9.5 0.038 Stoves and ovens 11 89 Microwave ovens 8 93 12.2 0.043 0 25 0.015 0 75 Window air conditioners 5.3 37 Televisions 0.128 14.1 62 0 58 14.7 Radios 0.023 42 0 Stereos 0 41 15.4 0.075 59 Rugs 39 0.062 7.0 58 3 77 23 Window coverings 0.060 2.5 0 0 11.3 0.011 Clocks 56 44 Lamps and lights 0 62 10.5 0.049 38 Telephones 67 0.012 4.8 33 0 86 1 13 9.8 0.085 Lawn and garden equipment 64 0 36 Power tools 0.035 3.2 a 10.6 2 22 0.042 Vacuums, sewing machines 76 a 56 0 44 Small kitchen appliances, heaters 0.072 8.0 25 0 0.007 Hard flooring 2.9 75 0 7.4 0.122 Office furniture 69 31 Hand tools 44 56 3.7 0.026 0 9 0.193 39 51 Men’s suits 4.7 22 Men’s coats and sportscoats 0.119 12.0 67 11 6 5.9 Men’s and boys’ sleepwear 0.044 94 0 Men’s and boys’ sweaters 16 54 20.2 0.043 30 Men’s pants 10 0.212 5.4 79 11 76 24 Boys’ coats, suits, and sportscoats 0.035 22.0 0 18 26.3 0.192 Women’s and girls’ coats 27 56 Women’s and girls’ dresses 21 56 38.3 0.284 23 Women’s sweaters and vests 19 0.072 27.7 61 20 63 21 16 14.4 0.394 Women’s skirts and pants 75 7 19 Women’s and girls’ sportswear 0.086 29.1 24.9 82 18 Women’s sleepwear 0.068 0 57 0.168 Women’s suits 23 32.4 20 4 7.9 0.224 14 Men’s footwear 82 Boys’ and girls’ footwear 0 83 15.1 0.154 17 3 18 Women’s footwear 0.341 11.4 79 0 0.078 Watches 8.7 77 23 66 31 7.5 0.323 Jewelry 3 60 0 40 Luggage 0.035 10.9 30 Cars 35 4.811 16.5 35 23 1.120 Trucks 50 15.6 26 83 17 2.5 0.256 Tires 0 50 13 37 Eyeglasses and contacts 0.335 2.9 47 Sports and exercise equipment 51 0.210 8.1 2 10.3 0 22 0.181 Bicycles 78

14 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH 1019 VOL. 91 NO. 4 ABLE T —Continued. 3 (2) (3) (4) (5) (1) 1997 item- Percent substitution Percent of Percent of weight in Percent of substitutions substitutions December rate (in substitutions “direct” “linked” percent) “comparable” Good 1997 CPI a 55 Camping/fishing/hunting equipment 0.046 7.5 2 43 Winter/water sports equipment 8.3 45 2 53 0.163 0.001 Playground equipment 37.5 100 0 0 0.062 Musical instruments 5.3 56 4 41 0.048 0 26 Photographic equipment 74 7.0 Personal care appliances 8.9 71 0 29 0.011 a Calculators, typewriters 0 0 100 2.4 0.004 11.2 61 5 34 Mean 8.4 Median 61 0 30 22 10 22 8.2 Standard deviation 100 50 Maximum 38.3 100 0 2.4 0 0 Minimum Weighted mean 13.8 46 22 32 25 48 3.8 ALL price quotes in the CPI 27 Nonresidential price quotes 3.3 58 13 29 8 Nonresidential, nonvehicle 29 3.0 63 Item-substitution rate: percentage of price quotes for which a substitute replaced the previous month’s item. (Because Notes: these are monthly, the fraction of items with some replacement during the year is much higher.) “Comparable” substitutions: the replacement item is treated as the same as the previous month’s item for pricing purposes; thus no quality adjustment is made. “Direct” quality adjustments: the price of the replacement item is divided by a measure of its quality relative to the previous month’s item. Quality is measured using hedonics or the manufacturer’s estimate of the cost of producing the replacement item relative to the previous item (gross of a markup). The “Link” method: the price of the replacement item is multiplied by the gross inflation rate of other items in the same category and divided by the ratio of its price to the price of the previous month’s item. a Four pairs of categories had to be combined because of lack of finer BLS data. sectional choices among goods, could existing ative to the BLS rate of inflation for nondura- bles. To be comparable to our construction of very well predict the rate at which consumers shift the unit prices, the BLS inflation rates are also into improved models, not just switch among in- based on a three-year moving average of defla- cumbent models. First, retailers may be upgrading tors. Across the 66 goods the correlation be- the quality of all models they sell (and manufac- tween the unit price changes in column (1) and turers all models they produce) in response to the BLS price changes in column (3) is 0.48. rising demand for quality. Demand for quality Figure 2 plots each good’s rate of BLS price should be rising faster where the quality slopes are inflation versus its rate of unit price inflation. steeper—the first term in equation (8). New mod- Microwaves are clearly an outlier in terms of els may appear all along the price– quality menu, both inflation rates. Dropping microwaves from not just at the very top. Second, if all qualities the sample reduces the correlation between the become cheaper, then our quality slopes interacted with the change in quality-adjusted prices should two inflation rates from 0.48 to 0.33. BLS deflators are not available for the full predict where quality upgrading will be rapid. 1980 –1996 period for all 66 goods. For 26 goods This is the second term in (8). In Table 2 we compare the BLS measures of the BLS sample period is shorter than 1980 price inflation to our constructed measures of through 1996 (see the notes to Table 2). Column unit price inflation good by good. The rate of (2) provides the rate of unit price inflation for the time period that the BLS price deflator is avail- unit price inflation, as discussed earlier, appears able. Comparing columns (2) and (3), the BLS in column (1). The rate of BLS inflation appears price inflation rates are systematically lower than in column (3). The BLS rates of inflation, like our unit price inflation rates, are expressed rel- the unit price inflation rates, presumably reflecting

15 SEPTEMBER 2001 1020 THE AMERICAN ECONOMIC REVIEW quality Engel curve for each of the 66 goods. u The estimate of a good’s quality slope is i based on how the unit price that a household pays for a good, say televisions, is related to a household’s total nondurable consumption. Generalizing (4) from the consumer’s prob- lem to include measurement error and ignoring terms that do not vary across households, we have 5 n ˆ c ln , u 1 ln ˆ x ln (15) 1 ́ i iht ht iht iht where x ˆ ˆ c iht ht 2. I ACH OF THE E ATES FOR R NFLATION IGURE F 66 G OODS u ln 2 ln 5 ́ D D S S i iht c x ht iht BLS adjustments for quality improvements. The x and last column in Table 2 reports the rate of inflation ˆ ˆ denote a household’s reported and c iht ht 15 in unit prices minus the rate of BLS inflation. We x values for The distinction be- c and . iht ht x tween the reported and true values for calculate this difference using unit price inflation and iht over the same period that the BLS inflation rate is c . In arriving at contributes the error term ́ iht ht (15) we are assuming that households face the available [i.e., we calculate it as column (2) minus column (3)]. Weighting by CPI shares, the mean z same quality-adjusted prices . In pooling it difference across the 66 goods is 1.46 percent cross sections of households from different faster inflation in unit prices than in BLS prices. years of the CEX, we add dummies for year, An interpretation of this is that the BLS incorpo- region, and city (versus rural) to control for rates quality growth of 1.46 percent per year on likely differences in prices across time and 14 Note that these quality average for these goods. space. In addition to heterogeneity in c ,we ht allow for heterogeneity in the household’s pref- adjustments are partly “active” (involving item erence for each good by including a number of substitution procedures), but may be mostly “pas- sive” (inflation in unit prices from consumers household characteristics as control variables. The household characteristics are number of upgrading among existing goods does not contam- persons and number of children in the house- inate BLS fixed-weight inflation). Moulton and hold, average age of the household head and Moses (1997) report that “active” quality adjust- ments amounted to between 0.28 and 0.44 percent that age squared, and dummy variables for sin- in 1995. If this is typical of active adjustments gle male-headed households and for single over 1980 –1996 for our set of durables, then most female-headed households. We interpret these 16 For five of variables as shifting n of our 1.46 percent estimate would stem from in (15). iht passive BLS quality adjustments. 15 Conditional on a household reporting more than one III. Estimating Quality Engel Curves from purchase of a good, we average the expenditures to arrive at Cross Sections of Households an average unit price. 16 Additional variation in this preference parameter is We employ CEX cross sections of house- another potential source of error in (15). Selection of house- holds for 1980 to 1996 to estimate a separate hold h into the sample of purchasers of good i based on the household’s value of n could bias the estimates of u ih i downward. If poorer households are less likely to buy a 14 good, then poorer households in the sample of purchasers The unit price and BLS price inflation rates also differ because the BLS weights (on outlets and on goods within will be those with a high preference for the good. It is not CEX categories) move only gradually, whereas current estimates of clear how this selection will bias the u relative i weights are embedded in average unit prices. across goods, which is central to our constructed instrumen-

16 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1021 the goods (carpeting, curtains and drapes, win- trend term was not significantly different from dow coverings, lamps and lights, and hardwood zero at the 0.05 level for 57 of the 66 goods (two were significantly negative; seven were flooring), we are concerned that richer house- significantly positive). So typically we cannot holds buy a larger size or quantity, as well as higher quality. For these goods we also control reject stability of the quality slopes. for the number of rooms in the household’s We also explored the appropriateness of the home. loglinear formulation. We compared our log c ˆ linear estimates to nonparametric (kernel) esti- We define in (15) to be a household’s ht total nondurable consumption. Our measure mates, and found no distinct patterns of convex- ity or concavity, nor any distinct patterns of of nondurables is narrower than that in the 17 Figures 3 and 4 illustrate by floors or ceilings. National Income and Product Accounts, in that we exclude clothing and footwear from comparing the linear and nonparametric Engel nondurables. To the extent that there is mea- curves for cars and vacuums, respectively. Cars are a high-expenditure good among those with surement error in a household’s response for c the steepest quality slopes. Vacuums are a high- , as allowed for in (15), an ordinary least ht expenditure good among those with the flattest u squares (OLS) estimate of will be biased i quality slopes. The linear estimates track the toward zero. For this reason we instrument for nonparametric estimates quite well, especially c ˆ - as follows. For each household we sepa ht 1 0.5, 2 over the ( rate spending on nondurables in the first and 0.5) range containing 88 per- second interview quarters from those in the cent of the log consumption observations. In estimating the quality slopes we have as- third and fourth interview quarters. We treat ˆ sumed that the higher unit prices paid by richer c as nondurable consumption measured for ht the latter two quarters, then instrument for households reflect the purchase of higher- quality versions of goods, not higher price this consumption with the household’s mea- markups conditional on quality. Might richer sured consumption in the first two interview households pay higher markups than poorer quarters. Consistent with there being mea- households do for the same quality of good? For surement error, the coefficient obtained by instrumenting is modestly higher for each cars, at least, this does not appear to be the case. Pinelopi Koujianou Goldberg (1996) finds no good than the coefficient obtained with OLS. correlation between the price a household pays Results for the quality Engel curves with estimation by two-stage least squares are pre- for a particular car model and the household’s sented in the first column of Table 4. Standard income, financial assets, education, or occupa- errors are in parentheses. The elasticities vary tion. We touch on this issue again below, but considerably. The steepest quality Engel curves note that markups would have to covary a lot are for jewelry, window coverings, rugs, and with nondurable consumption to explain Table cars. A 1-percent increase in nondurable spend- 4. If Household A has twice the nondurable consumption of Household B, then Household ing is associated with about a 1-percent increase A typically pays about 76 percent more for a in purchase price for these goods. At the other consumer durable. This is substantially larger extreme, prices for microwave ovens, sewing machines, vacuums, and lawn and garden level than most estimates of the of markups. We next compare our quality slopes to the equipment each exhibit unit price elasticities with respect to total nondurables of 0.25 or less. steepness of the overall Engel curve for each good. In the second column of Table 4 we We tested the stability of the quality slopes report quantity Engel curves constructed as fol- c over time by adding a variable interacting ln lows. For each good a household’s quantity of with a linear time trend. The coefficient on this 17 tal variable. Such selection, if important, will also occur For the kernel estimation we used the default in Eviews: bandwidth an Epanechnikov kernel with over time. As economy-wide income and consumption rise, p 0.15 ), local linear regression, linear (max ln c 2 min ln c the amount of quality upgrading in the average purchase binning, and 100 gridpoints. After estimation, but before price of a good will, similar to the cross-section pattern, be plotting, we trimmed the top and bottom 1 percent of the biased down by the entry into the markets of consumers ln c observations. with a relatively low preference for the good.

17 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1022 NGEL C URVE S LOPES T 4—E ABLE (3) (1) (2) Quality/ (quality 1 quantity) Quantity (in percent) Quality Good 0.75 (0.08) 0.61 (0.05) Carpeting 55 70 0.39 (0.03) 0.93 (0.04) Curtains and drapes Mattress and springs 0.62 (0.04) 0.65 (0.04) 49 Bedroom furniture 48 0.75 (0.04) 0.70 (0.05) 0.53 (0.04) 59 Sofas 0.76 (0.04) 54 0.63 (0.03) 0.75 (0.04) Living room furniture 0.67 (0.04) Kitchen/dining room furniture 0.84 (0.06) 56 Baby furniture and equipment 0.46 (0.04) 0.45 (0.05) 50 1.00 (0.04) 48 Outdoor furniture 0.93 (0.05) Refrigerators and freezers 0.46 (0.04) 0.35 (0.04) 57 0.28 (0.04) 0.37 (0.05) Clothes washers 43 0.67 (0.06) 0.32 (0.05) Clothes dryers 32 Stoves and ovens 0.41 (0.06) 0.48 (0.06) 46 Microwave ovens 0.16 (0.03) 0.53 (0.05) 23 Window air conditioners 0.31 (0.07) 46 0.26 (0.08) 0.50 (0.03) 0.41 (0.03) Televisions 45 Radios 37 0.37 (0.03) 0.65 (0.03) Stereos 0.34 (0.04) 1.05 (0.04) 25 1.07 (0.05) 56 0.85 (0.04) Rugs 0.56 (0.04) 1.11 (0.06) Window coverings 66 Clocks 0.74 (0.04) 0.50 (0.04) 60 0.80 (0.03) Lamps and lights 0.81 (0.04) 50 Telephones 0.73 (0.03) 45 0.59 (0.03) 30 Lawn and garden equipment 0.25 (0.05) 0.57 (0.03) Power tools 32 0.29 (0.04) 0.60 (0.04) 25 0.24 (0.04) Vacuums 0.75 (0.04) 35 Sewing machines 0.19 (0.10) 0.36 (0.08) 38 0.39 (0.02) Small kitchen appliances 0.65 (0.02) 60 0.28 (0.04) 0.41 (0.03) Heaters 68 Hard flooring 0.30 (0.11) 0.64 (0.15) 39 0.71 (0.07) Office furniture 1.11 (0.06) Hand tools 0.55 (0.03) 0.58 (0.03) 49 0.68 (0.02) 31 Men’s suits 1.52 (0.03) 0.61 (0.02) 1.24 (0.02) 33 Men’s coats and sportscoats 0.97 (0.03) 0.37 (0.02) Men’s and boys’ sleepwear 27 0.46 (0.01) 1.13 (0.02) 29 Men’s and boys’ sweaters 0.45 (0.01) Men’s pants 39 0.71 (0.01) 41 Boys’ coats, suits, and sportscoats 0.48 (0.02) 0.68 (0.03) Women’s and girls’ coats 0.57 (0.01) 1.08 (0.02) 34 Women’s and girls’ dresses 0.96 (0.01) 41 0.67 (0.01) Women’s sweaters and vests 0.50 (0.01) 1.11 (0.02) 31 37 0.89 (0.01) 0.52 (0.01) Women’s skirts and pants Women’s and girls’ sportswear 0.47 (0.01) 1.28 (0.02) 27 0.97 (0.02) Women’s sleepwear 0.44 (0.01) 31 33 0.72 (0.02) Women’s suits 1.44 (0.03) 48 Men’s footwear 0.57 (0.01) 0.52 (0.01) 54 0.50 (0.01) 0.43 (0.02) Boys’ and girls’ footwear 0.70 (0.01) 0.62 (0.01) Women’s footwear 47 0.68 (0.02) Watches 49 0.70 (0.02) 52 1.06 (0.02) 1.13 (0.02) Jewelry Luggage 0.90 (0.04) 1.54 (0.04) 37 0.39 (0.02) Cars 0.94 (0.03) 71 74 0.33 (0.04) Trucks 0.93 (0.06) 0.42 (0.02) 0.67 (0.02) 38 Tires Eyeglasses and contacts 0.27 (0.02) 0.75 (0.02) 26 1.30 (0.03) 31 0.59 (0.03) Sports and exercise equipment

18 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH 1023 VOL. 91 NO. 4 ABLE 4 —Continued. T (2) (1) (3) Quality/ (quality 1 quantity) Quantity (in percent) Quality Good 39 0.67 (0.04) 0.43 (0.05) Bicycles Camping equipment 34 0.95 (0.06) 0.50 (0.06) Fishing and hunting equipment 53 0.59 (0.04) 0.66 (0.04) 0.81 (0.05) 1.45 (0.04) 36 Winter/water sports equipment 0.68 (0.13) Playground equipment 49 0.71 (0.08) Musical instruments 0.37 (0.07) 0.90 (0.05) 29 Photographic equipment 0.65 (0.04) 0.97 (0.04) 40 28 Personal care appliances 0.34 (0.02) 0.90 (0.03) 30 0.81 (0.04) 0.35 (0.04) Calculators 46 0.68 (0.07) 0.57 (0.09) Typewriters 0.76 43 0.57 Mean 0.69 Median 0.54 41 Standard deviation 0.23 0.31 13 1.13 Maximum 1.54 74 Minimum 0.16 0.28 23 0.62 0.76 Weighted mean 56 Notes: Sample: Cross sections of households in the 1980 –1996 U.S. Consumer Expenditure Surveys. (1982–1996 for calculators, and 1983–1996 for telephones and typewriters.) Observations: 65,189 household-years for the Quantity regres- sions. For the Quality regressions, observations are household-years with purchases of the good. Thus the number of observations varies by good for the Quality regressions. See Table 1 for the number of observations for each good. The weighted mean is calculated using the CPI shares in December 1997. Across the 66 goods in the table, the correlation between the Quality and Quantity slopes is 0.20. The estimates in Table 4 show that the quan- purchases of the good is regressed on ln c as well as the time and household control variables tity Engel curves differ sharply across goods. employed in estimating the quality Engel All goods display elasticities of at least 0.28, curves: and 14 goods display elasticities greater than 1. The final column of Table 4 presents the size of ˆ # the quality Engel curve relative to the sum of 5 (16) quantity Engel curve slope ! / V V ~ i iht responses in quality and quantity (i.e., relative to the overall Engel curve that incorporates how ln ˆ c p error term. 1 ht both quality and quantity increase as nondura- ble consumption rises). The share accounted for As with the quality Engel curves, in estimating by the quality Engel curve ranges from a low of 23 percent for microwaves to a high of 74 the quantity Engel curves we instrument for percent for trucks. On average the quality re- nondurable consumption in quarters 3 and 4 sponse to nondurable consumption is actually with nondurable consumption in quarters 1 and 2. The sample here, however, is the full sample of 65,189 households, not just those purchasing the good. So that the regression response in , rather than the household’s own expenditure on the good i quantity can be interpreted as an elasticity, in good, which in many cases is zero. (16) we divide a household’s purchase quantity We are not interpreting the estimates of the quantity i of good by the mean purchase quantity for Engel curve slopes in terms of structural parameters. We 18 present these estimates as one benchmark for judging the i in the sample. good magnitudes of the estimated quality slopes. Structural inter- pretation of the quantity Engel slopes is complicated, for one, by the fact that we observe only expenditures rather 18 than stocks for the goods. This is discussed in detail in Bils The percentage response in a household’s expenditure is expressed relative to average household expenditure on and Klenow (1998), Section III.

19 SEPTEMBER 2001 1024 THE AMERICAN ECONOMIC REVIEW ACUUMS V URVE FOR C NGEL E UALITY 4. Q F F ARS C URVE FOR C NGEL E UALITY 3. Q IGURE IGURE more important in magnitude than the quantity A fixed additive bias in the slopes (say from a response: when weighted by expenditures, the constant elasticity of the durables markup with average share accounted for by the quality En- respect to nondurable consumption) would have ˆ , or even on the first-stage coef- m gel curve is 56 percent. no effect on Although we take Table 4 as supportive of an ficient from regressing unit price growth on the quality slopes. Proportional bias in the slopes important role for quality upgrading in growth, would bias the first-stage coefficient, but would we caution readers (especially potential calibra- tors) against a literal interpretation. The quality have no effect on the second-stage estimation of m slopes could be systematically biased upward or . Finally, differential bias in the quality slopes that was uncorrelated with true inflation would downward. For instance, they might be biased reduce the first-stage fit and hence the precision upward if richer households tend to pay higher of the second-stage estimation, but would not markups, controlling for quality, or if richer households lump more purchases together in the ˆ. m bias CEX. Nonseparabilities of durables and nondu- IV. Estimating Quality Changes rables consumption could bias the slopes in either direction, as could selection bias. The A. Quality Engel Curves and Unit slopes could be biased downward if measure- ment error in nondurable consumption remains Price Inflation even after instrumenting with lags. Moreover, part of the quality Engel curve could be misat- We first ask if a good that exhibits a large unit tributed to the quantity Engel curve if richer price response to consumption cross sectionally households replace their durables with greater (a steep quality slope) also displays a faster increase in unit prices over time. The answer, it frequency. By replacing more frequently richer turns out, is yes. We then estimate to what households may have better, less-depreciated extent these predictable, quality-induced varia- durables on average. This would not be cap- tured in the unit price they pay, and therefore tions in unit price inflation contaminate BLS 19 would not show up in our quality slopes. estimates of a good’s price inflation. Fortunately, our IV estimation of m (the share There is a strong positive relation, as conjec- of quality growth that goes unmeasured) is ro- tured, between the slope of a good’s quality bust to many forms of bias in the quality slopes. Engel curve and its rate of unit price inflation. The correlation equals 0.51, suggesting the quality slope is a highly relevant instrument. Figure 5 plots the rates of unit price inflation 19 For a number of our goods this could be investigated against the quality slopes for the sample of 66 systematically using the durable goods inventory portion of the CEX. goods. Microwave ovens are an outlier because

20 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH 1025 VOL. 91 NO. 4 elry, rugs, and window coverings from the sam- ple. These goods exhibit an estimate of u two i or more standard deviations from the mean value of 0.57. Row 3 shows that the resulting coefficient and t -statistic are virtually the same as with the full sample. Because cars and trucks are outliers in terms of their CPI weight in the regressions, together receiving 48-percent weight (39 percent for cars; 9 percent for trucks), in row 4 of Table 5 we report results omitting them. The results change only modestly (coefficient 3.21 percent -statistic versus 4.24 percent in the full sample, t 5.2 versus 5.8), so they do not hinge on the vehicle categories and their large weighting. Running unweighted least squares on the full U UALITY RICE P NIT I LOPES AND F IGURE 5. Q NFLATION S sample yields similar results: a coefficient of ATES FOR OODS R 66 G 4.11 percent with a t -statistic of 4.7. Finally, one could argue that the apparel categories (16 of their very low rate of unit price increase. A of the 66 categories, with 21 percent of the CPI very strong positive relation remains, however, weight in the regressions) are not independent observations, so row 5 uses only the 50 nonap- if we remove microwaves, with the correlation parel categories. Although the coefficient is equaling 0.48. Recall that (9) predicts a faster unit price modestly lower than the baseline estimate (3.66 percent versus 4.24 percent) and the u -statistic is t inflation rate the steeper the quality slope .In i lower, reflecting the smaller sample and coeffi- Table 5 we report results from weighted least- cient (3.4 versus 5.8), the coefficient remains squares regressions (with the weights equaling highly significant. December 1997 CPI shares). The dependent variable is average unit price inflation over Could the coefficient from this regression (e.g., the 4.24-percent baseline estimate) 1980 –1996 for good i , and the independent variable is the quality slope estimated for good plausibly reflect consumers upgrading quality from 1980 –1996 cross sections of the CEX. i faster for goods with steeper quality slopes? Hence there is one observation per consumer From (9), we anticipate a coefficient on the 20 D s durable category for 66 observations in the full c D Growth in 2 . z quality slope equal to c sample. As shown in row 1 of Table 5, the D real per capita nondurables consumption hypothesis that unit price inflation is unrelated averaged 1.26 percent per year from 1981 to 1995. Relative to the price of nondurables, u to -statistic of 5.8. t is easily rejected with a i BLS prices for our set of durable goods fell by The coefficient implies that a unit increase in 0.82 percent per year on average (weighting the quality slope (roughly the difference be- tween the steepest and flattest slopes among the goods by their CPI shares). For illustration equals 1 (utility is logarithmic 66 goods) is associated with 4.24 percent faster s suppose that D c 2 in nondurable consumption). Then unit price inflation over 1980 –1996. z D s To check robustness of this first-stage regres- , the impact of the quality slope on infla- tion in unit prices, should be 2.08 percent sion we reestimated after eliminating micro- waves and trucks from the sample. These were [ 5 2 0.82 percent)]. ( 2 1.26 percent the only goods with rates of unit price inflation The preceding calculation assumes, however, that there is no unmeasured quality growth for two or more standard deviations from the mean of 2 0.44 percent per year. Row 2 of Table 5 shows that, excluding these two goods, the u coefficient on falls slightly from 4.24 to 4.13 20 i and z D is uncorrelated with both u This assumes that i i t percent and the -statistic rises considerably to D n is required for validity of . Orthogonality of u and D z i i i u 12.1. We also reestimated after eliminating jew- n as an instrument, but orthogonality of u and D is not. i i i

21 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1026 P RICES NIT U HANGES IN C REDICTING 5—P ABLE T Number of u Coefficient on i 2 (percent) observations R Adjusted Weighted least-squares regressions Full sample of goods 0.93 66 4.24 (0.72) 5 5.8 t 64 0.98 4.13 D x Minus 2 1 SD extremes i (excludes microwave ovens and trucks) (0.34) t 5 12.1 4.25 Minus 2 1 SD 0.93 63 u extremes i (excludes jewelry, rugs, and window (0.75) coverings) 5.7 5 t Minus CPI weight extremes 0.31 64 3.21 (0.62) (excludes cars and trucks) 5 t 5.2 0.93 3.66 50 Minus apparel (excludes the 16 clothes (1.06) and shoes categories) t 5 3.4 x Notes: The weighted least squares weights are equal to December 1997 CPI shares. The dependent variable is D (percent i unit price growth for good i ) averaged over 1980 –1996. The regressor is u , the quality slope for good i . According to i equation (9) in the text, the coefficient on u should equal D c 2 z s D . This regression is the first-stage regression for the i instrumental variables estimation that follows in Table 6. 21 durables. Filling the gap between 2.08 percent egy requires enough change in the level of non- and the 4.24-percent coefficient in Table 5 re- durable consumption or in the relative price of quires unmeasured quality growth of 2.16 per- our durables. If 1980 –1996 had been a period D s cent per year on average across our durable c D 2 was stagnant, the quality z over which slopes would have had no predictive power. But 1). This is in line with the degree 5 s goods (if 2 of unmeasured quality growth we estimate for - R as the adjusted values in Table 5 demon our goods below (2.2 to 2.4 percent per year). strate, the first-stage fit is ample, consistent with We also note that, by multiplying the coeffi- evidence that nondurable consumption grew and durables prices fell. cient in this regression by the average value of We assume that a good with a steep quality u of 0.76, we arrive at an estimate of the i average rate of quality upgrading for our goods. slope exhibits fast unit price inflation because of For the coefficient of 4.24 percent, the average fast quality growth, not fast true inflation. A implied quality growth is 3.2 percent per year. good with a steep quality slope will also typi- This is reasonably close to what we estimate cally exhibit a steep overall (quantity plus qual- below (3.7 to 3.8 percent). ity) Engel curve. For this reason, the demand for resources to produce this good should be rising. Table 5 is the first-stage regression for the If the industry exhibits constant returns to scale m second-stage estimation of (see Table 6 be- then this will not affect the price per unit of low). It is important to emphasize that the first- quality for the good. If returns to scale are not stage prediction of time-series unit price inflation with cross-sectional quality slopes constant, however, then steepness of the overall Engel curve will affect the good’s price per unit need not have worked. As (9) shows, our strat- of quality. One test of our assumption of con- stant returns is to see how price responds to a 21 The discussion also assumes no unmeasured quality good’s quantity Engel curve (those we reported growth for nondurables. However, each percent of un- in Table 4), because a steep quantity Engel measured quality growth in nondurables understates both curve also predicts rising demand for the prod- by 1 percent. Thus, for D c and D z s 1, it has precisely 5 D offsetting effects on the two terms in ( D c 2 s ). uct over time. Repeating the first row regres- z

22 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1027 sion, now including the good’s quantity Engel We conclude that a good’s quality slope ro- bustly predicts its unit price inflation rate. curve, yields an insignificant coefficient on the 0.89 percent (stan- 2 quantity Engel curve of dard error 0.66), and one with the opposite sign B. Quality Engel Curves and predicted by upward-sloping marginal cost. The BLS Price Inflation coefficient on the quality Engel curve falls, but , the share remains highly significant at 3.50 percent with a m We are now prepared to estimate t of quality growth that gets mismeasured as in- -statistic of 3.9. standard error of 0.91 and Related, we re-estimated adding the change is identified by com- m flation. Our estimate of in the share of CEX households buying each bining (7), (9), and (13) with conditions that the D residual z good as a control variable. More households be orthogonal to our instruments u i i should be buying goods whose quality-adjusted and ( u - ) u . Estimation is by General z 2 D i i ized Method of Moments (GMM) and the relative price has fallen. Including this variable , actually increases the coefficient on the quality results appear in Table 6. We first estimate m Engel curve from its baseline value of 4.24 employing only the quality slope as an instru- ment (row 1). We clearly reject the hypothesis percent to 4.78 percent (standard error 0.37, that 5 0( t -statistic 4.9). Moreover, the m t -statistic 13.0). m is sizable, equaling 0.618 with a estimate of Finally, we investigated whether aspects of the producing industries suggest goods with standard error of 0.125. This means that BLS steeper quality slopes might have increasing prices rise by 61.8 percent as much as do unit prices in response to quality upgrading pre- quality-adjusted prices, which would violate identifying condition (10). True inflation might dicted by a good’s quality slope. If the BLS be faster for labor-intensive industries (those quality adjustments, which average 1.46 percent with low capital–labor ratios or high labor per year across our goods, miss 61.8 percent of shares in value added) or industries with rapid quality growth, then true quality growth equals growth in wages or materials prices and slow 2 1.46 percent/(1 5 3.82 percent per year [ 0.618)]. The quality bias in BLS inflation rates growth in TFP. Using four-digit manufacturing industries in the NBER Productivity Database, for our goods would then be 2.4 percent per we examined how these industry characteristics year (3.82 percent minus the BLS quality ad- justments of 1.46 percent). correlated with the quality slope of the good Table 6, row 2, presents results adding as an produced. We averaged over 1980 –1996 and weighted each industry by its CPI share. We instrument the interaction between ( u u ) 2 i z and D found mostly small and insignificant correla- m . The estimate of is modestly reduced i tions of these industry variables with the quality to 0.601 with a standard error of 0.119 and a -statistic of 5.0. The implied average growth in t slopes across the 66 goods. The only exceptions quality across our 66 goods is then 3.7 percent were with the equipment capital-to-labor ratio -value 0.11) and with 1 [ 5 p 1.46 percent/(1 2 0.20 and 0.601)]. This exceeds the (correlation 0.43, actual 1.46 percent BLS adjustment by 2.2 1 p TFP growth ( -value 0.0003). These percent per year, implying that BLS inflation for correlations suggest, if anything, that true infla- tion might be lower for goods with steeper our goods is biased upward by 2.2 percent per quality slopes. The significantly more rapid TFP year. could be overstated if Our estimate of m growth is particularly suggestive, because we might have expected a negative correlation quality-adjusted price changes are positively given our finding (below) that inflation is more correlated with our quality slopes. That is, if goods with steep quality Engel curves happen to overstated for goods with steeper quality 22 slopes. percent per year relative to unit prices for vacuums, accu- 22 mulating to 60-percent greater inflation for cars over the It is also difficult to explain the differences in rates of unit price inflation predicted in Table 5 on the basis of period of 1980 to 1996. Markups would need to have changing price markups over marginal cost. For instance, increased markedly for goods with steep quality slopes, the difference in quality slopes between cars and vacuums relative to other goods, to play an important part in such of 0.7 predicts unit prices for cars would increase by 3 large relative changes in unit prices.

23 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1028 NFLATION IAS B I AND , ROWTH G UALITY ,Q m STIMATES OF 6—E ABLE T Upward Average Adjusted inflation bias quality growth Instrument set 2 R (percent per year) m 2 (percent per year) u 3.8 2.4 0.56 0.618 i (0.125) 5 4.9 t 0.57 2.2 u 3.7 0.601 2 u ) ,( z u D i i i (0.119) t 5 5.0 the fraction of quality growth that shows up as inflation in the BLS price The number of observations 5 66. m 5 Notes: u deflators. i the growth rate of the quality-adjusted relative price of good 5 the quality slope for good . D (relative i z 5 i i p to the price of nondurable consumption). Estimation: The estimating equation is D 2 m D x 5 1 (1 z m ) z D z . This is i i i equation (7) in the text. Here is estimated by exploiting m m is estimated by GMM using the instruments listed above. That is, z D the orthogonality of to the instruments given. Average quality growth: The difference between the unit price inflation rates i x D is an estimate of the BLS’s quality adjustments. Across our 66 goods, these quality and the BLS inflation rates D p i i adjustments averaged 1.46 percent per year (when the goods are weighted by their 1997 CPI share). Thus if the BLS adjustments are capturing only (1 2 ) of total quality growth, total quality growth must be 1.46/(1 ). This is equation m m 2 (13) in the text. Upward inflation bias: The BLS misses the fraction m of total quality growth, which equals 1.46 z m /(1 2 ). This is equation (14) in the text. m other estimates in the literature? The Boskin have slower rates of cost-reducing technologi- cal progress or face faster growth in factor Commission (Boskin et al., 1996) estimated prices, then their prices will be rising for a quality bias of 0.6 percent per year for the overall CPI, but 1.0 percent per year for the reason in addition to quality upgrading. We presented evidence in the preceding subsection consumer durable subcomponent (our calcula- that this is not the case. As one additional effort tion from the breakdown in their Table 2). Gor- don (1990) estimated that the BLS price index to address this possibility, we reestimated ex- cluding goods with rates of BLS price inflation for consumer durables was overstated by at least 1.5 percent per year from 1947 to 1983, and at more than two standard deviations away from the mean of 1.33 percent per year. This elim- least 1.0 percent per year from 1973–1983. Gor- 2 inated five goods from the sample (microwaves, don considered his estimates lower bounds for TVs, radios, telephones, and luggage) and low- at least two reasons. First, Gordon stressed that to 0.477 (standard error m ered the estimate of BLS techniques also fail to account for im- -statistic 4.6). This would imply an in- 0.104, proved quality from greater durability (e.g., of t flation bias of 1.3 percent per year, versus the 2.2 percent implied by estimation of with the m 23 full sample of goods. dow coverings) lowered the estimate to 0.568 (0.117, How do our estimates of bias compare to 4.9). We recalculated D x based on the periods BLS prices are i 23 We conducted a number of other robustness checks: available, rather than using the entire 1980 to 1996 period. - D Using this alternative measure of x to construct the in i u strument ( 2 u ) D z had very little effect. The estimate of (i) Weighting goods equally lifted the estimate of m to i i became 0.622 (0.122, 5.1). Using this alternative measure m t 0.812 (standard error 0.199, -statistic 4.1). in the first-stage regression (9), as well as in constructing the (ii) Excluding cars and trucks (the CPI weight extremes) u instrument ( u 2 estimate of 0.657 (0.163, m , led to a z D ) i i resulted in a estimate of 0.884 (0.222, 4.0). m 4.0). (iii) Excluding the 16 clothing and shoes categories low- Finally, we also tested whether the m coefficient system- to 0.561 (0.145, 3.9). ered the estimate of m atically differs in size for those goods for which the BLS (iv) Excluding the unit price growth extremes (micro- implicitly makes a large quality adjustment (goods with a waves, trucks) boosted the estimate of m to 0.680 large value in the final column of Table 2) or goods that (0.153, 4.5). constitute larger shares in consumer spending. We found no (v) Excluding quality slope extremes (jewelry, rugs, win- significant interactions.

24 BILS AND KLENOW: QUANTIFYING QUALITY GROWTH VOL. 91 NO. 4 1029 automobile tires) and increased energy effi- error, our estimates imply that at least one-third ciency (e.g., of appliances). Second, Gordon of quality growth flowed through into measured assumed zero bias in the consumer durables that inflation, biasing consumer durables inflation by he did not examine (about one-half of expendi- at least 0.8 percent per year over 1980 –1996. We should add that our approach does not tures on durables). yield good-by-good or period-by-period esti- To summarize, differences in quality slopes successfully predict differences in unit price mates of quality growth and quality bias. The approach provides an overall diagnostic on the inflation rates. These differences pass through extent of quality bias in official inflation rates into differential rates of BLS price inflation. for a set of goods. Yet a strength of our ap- Our preferred estimate of m is about 0.60, which proach relative to using hedonics is that our implies about 2.2-percent upward bias in BLS approach does not require detailed information inflation for our consumer durables because of on the attributes of goods. Our approach re- failure to fully net out quality growth. As a cautionary note, although we can reject the hy- quires data only on unit prices and on simple 5 0 with considerable confidence, pothesis of attributes of In this paper we have fo- buyers. m cused on the richness of buyers, but other at- our estimate of m is associated with a nontrivial tributes that are correlated with unit prices could standard error. The two standard error bands be used as well, such as age or household com- contain 0.363 and 0.849. This translates into a position (number of kids, number of workers, fairly wide confidence interval in assigning a etc.). With scanner data from supermarkets, de- particular number to unmeasured quality partment stores, and the like, data on unit prices growth. We can say, with greater confidence, that our estimates imply that at least one-third of could become accessible for a much wider set of goods than our 66 durable goods comprising 12 quality upgrading was mismeasured as inflation ( m 0.363, our point estimate minus two stan- percent of the CPI. One must be able, however, 5 24 to match these unit prices to buyer attributes. dard errors), and that this generated a bias of at least 0.8 percent per year. This would be asso- REFERENCES ciated with quality growth of 2.3 percent per year, only 1.5 percent per year of which was “On an Index Adelman, Irma and Griliches, Zvi. netted out by BLS adjustments. of Quality Change.” Journal of the American Statistical Association , September 1961, V. Conclusion (295), pp. 535– 48. 56 Aghion, Philippe and Howitt, Peter. “Growth and We estimated quality Engel curves for 66 , March Creative Destruction.” consumer durables from pooled cross sections Econometrica (2), pp. 323– 61. 60 1992, of households in the 1980 through 1996 U.S. Berry, Steven; Kortum, Samuel and Pakes, Ariel. Consumer Expenditure Surveys. We used their slopes to predict the speed of quality upgrading “Environmental Change and Hedonic Cost Functions for Automobiles.” National Bureau for the goods. Just as if households were as- of Economic Research (Cambridge, MA) cending their quality Engel curves over time, we found that the average price paid rose faster Working Paper No. 5746, September 1996. for goods with steeper quality Engel curves. Bils, Mark and Klenow, Peter J. “Using Con- sumer Theory to Test Competing Business BLS prices likewise increased more quickly for goods with steeper quality Engel curves, sug- Journal of Political Econ- Cycle Models.” omy (2), pp. 233– 61. 106 gesting the BLS did not fully net out the impact , April 1998, of quality upgrading on prices paid. We esti- Boskin, Michael J.; Dulberger, Ellen R.; Gordon, Robert J.; Griliches, Zvi and Jorgenson, Dale. mated quality growth of about 3.7 percent per year for our goods. We estimated that BLS “Toward a More Accurate Measure of the quality adjustments captured about 40 percent of this upgrading, with roughly 60 percent, or 2.2 percent per year, showing up as higher in- 24 Unit prices for some nondurables might be available flation rather than higher real growth. Even in- in the Diary Surveys of the CEX, which would contain the necessary buyer attributes. corporating alternative samples and sampling

25 SEPTEMBER 2001 THE AMERICAN ECONOMIC REVIEW 1030 Cost Of Living.” Final Report to the Senate . National Bureau of nomics of new goods Finance Committee, December 4, 1996. Economic Research Studies in Income and Wealth, Vol. 58. Chicago and London: Cutler, David M.; McClellan, Mark; Newhouse, Joseph P. and Remler, Dahlia. University of Chicago Press, 1997, pp. 29 – “Are Medical 66. Prices Declining?” National Bureau of Eco- nomic Research (Cambridge, MA) Working “Housing Depreciation Randolph, William C. and Aging Bias in the Consumer Price In- Paper No. 5750, September 1996. Goldberg, Pinelopi Koujianou. “Dealer Price Journal of Business and Economic Sta- dex.” , July 1988, (3), pp. 359 –71. 6 Discrimination in New Car Purchases: Evi- tistics “Hedonic Prices and Implicit Rosen, Sherwin. dence from the Consumer Expenditure Sur- Markets: Product Differentiation in Pure vey.” Journal of Political Economy , June Competition.” , Journal of Political Economy (3), pp. 622–54. 1996, 104 (1), pp. 34 –55. Gordon, Robert J. January–February, 1974, 82 The measurement of durable goods prices Shapiro, Matthew D. and Wilcox, David W. . Chicago: University of Chicago “Mis- measurement in the Consumer Price Index: Press, 1990. An Evaluation,” in Ben S. Bernanke and Griliches, Zvi. “Hedonic Price Indexes for Au- Julio J. Rotemberg, eds., NBER macroeco- tomobiles: An Econometric Analysis of nomics annual 1996 The price Quality Change,” in G. Stigler, ed., . Cambridge, MA: MIT , General Press, 1996, pp. 93–142. statistics of the federal government “Product Innovations, Series No. 73. New York: Columbia Univer- Trajtenberg, Manuel. Price Indices, and the (Mis)Measurement of sity Press for the National Bureau of Eco- Economic Performance.” National Bureau of nomic Research, 1961, pp. 137–96. “Generics and Griliches, Zvi and Cockburn, Iain. Economic Research (Cambridge, MA) Work- ing Paper No. 3261, February 1990. New Goods in Pharmaceutical Price Index- American Economic Review es.” , December Triplett, Jack E. “Measuring Consumption: The Post-1973 Slowdown and the Research 84 (5), pp. 1213–32. 1994, Federal Reserve Bank of St. Louis “Valuation of New Goods Issues.” Hausman, Jerry A. , May/June 1997, Review 79 (3), pp. 9 – Under Perfect and Imperfect Competition,” in T. F. Bresnahan and R. J. Gordon, eds., 42. Department of The economics of new goods U.S. Bureau of Labor Statistics. . National Bu- reau of Economic Research Studies in In- Labor. 1995 Consumer Expenditure Survey: Interview, EXPN, and Diary come and Wealth, Vol. 58. Chicago: . CD-ROM, Au- gust 17, 1998. University of Chicago Press, 1997, pp. 209 –37. . “Cellular Telephone, New Products, . 1995 Consumer Expenditure Survey: and the CPI.” Journal of Business and Eco- Interview, EXPN, and Diary . CD-ROM, June 16, 1999. , April 1999, nomic Statistics 17 (2), pp. 188 –94. “CPI Strata with Small Sam- Consumer Price “Selective Access.” . Moulton, Brent R. , ple Sizes.” U.S. Bureau of Labor Statistics, http://stats.bls.gov/cpihome.htm, Indexes Department of Labor, internal memorandum, November 2000. August 29, 1996. U.S. General Accounting Office, U.S. Congress. Moulton, Brent R. and Moses, Karin E. “Consumer Price Index: Impact of Commod- “Ad- dressing the Quality Change Issue in the Con- ity Analysts’ Decisionmaking Needs to Be , GAO/GGD-99-84, Brookings Papers on Letter Report sumer Price Index,” Assessed.” Economic Activity , 1997, (1), pp. 305– 66. June 15, 1999. Nordhaus, William D. “Do Real-Output and “Interview University of California, Berkeley. Consumer Expenditure Surveys.” EML Ar- Real-Wage Measures Capture Reality? The chive, ftp://elsa.berkeley.edu/pub/ices/, No- History of Lighting Suggests Not,” in T. F. The eco- vember 2000. Bresnahan and R. J. Gordon, eds.,

Related documents