Prices Rise Faster than They Fall
Author(s): Sam Peltzman
Source: The Journal of Political Economy, Vol. 108, No. 3 (Jun., 2000), pp. 466-502
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/3038267
Accessed: 09/09/2008 16:05
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=ucpress.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.
http://www.jstor.org
Prices Rise Faster than They Fall
Sam Peltzman
University of Chicago
Output prices tend to respond faster to input increases than to
decreases. This tendency is found in more than two of every three
markets examined. It is found as frequently in producer goods
markets as in consumer goods markets. In both kinds of markets
the asymmetric response to cost shocks is substantial and durable.
On average, the immediate response to a positive cost shock is at
least twice the response to a negative shock, and that difference is
sustained for at least five to eight months. Unlike past studies,
which documented similar asymmetries in selected markets (gasoline, agricultural products, etc.), this one uses large samples of diverse products: 77 consumer and 165 producer goods. Accordingly, the results suggest a gap in an essential part of economic
theory. As a start on filling this gap, the study finds no asymmetry
in the response of an individual decision maker (a supermarket
chain) to its costs, but it finds above-average asymmetry where a
cost shock is filtered through a fragmented wholesale distribution
system. It also finds a negative correlation between the degree of
asymmetry and input price volatility and no correlation with proxies for inventory costs, asymmetric menu costs of price changes,
and imperfect competition.
I want to thank Robert Book, Yuri Garbuzov,Seok-kyun Hur, Joseph Kaboski,
Sangheon Kim, YoungsikLim, Tim McKenna,William Pearson, Steve Tenn, Brian
Viard,and MaryBeth Wittekindfor research assistance.John Carlson,Anil Kashyap,
Joe Mattey,YoramWeiss,and seminar participantsat the Universityof Chicago and
anonymous referees provided valuable comments and suggestions. Mark Mrowiec
of Dominick's Supermarkets,Inc. guided me patiently through the nuances of the
Universityof Chicago GraduateSchool of Business'sDominick's database.Don Barnett, J. Upton Hudson (U.S. Steel), and F. M. Scherer (Scherer Engineers, Ltd.)
provided crucial help in matching inputs to outputs in the steel industry. Roger
Fisher (formerly of Amoco Chemicals) did the same for organic chemicals. The
usual caveat about errors applies with special force. Research support from grants
to the George J. Stigler Center for the Study of the Economy and the State by the
Lynde and HarryBradleyFoundation and the Sarah Scaife Foundation is acknowledged with gratitude.
Uournal of Political Economy, 2000, vol. 108, no. 3]
? 2000 by The University of Chicago. All rights reserved. 0022-3808/2000/10803-0006$02.50
466
PRICES
RISE
467
The Great Pork Gap: Hog Prices Have Plummeted.
Why Haven't Store Prices? [New YorkTimes,January 9,
1999]
I.
Introduction
This headline and accompanying story articulate a common view
among consumers. It is that the prices they pay promptly reflect input price increases but not decreases. And there is something
vaguely sinister about this asymmetry. This particular story is not half
obligatorily
over before the c (ollusion) -word is mentioned-and
pooh-poohed by an academic economist.
Apparently businesspeople have more varied views on this issue,
as perhaps befits their status as both buyers and sellers of goods.
Alan Blinder and his associates (Blinder 1994; Blinder et al. 1998)
asked them, "How much time normally elapses after a significant
increase (decrease) in cost before you raise (decrease) your prices?"
Their answers implied symmetric lags in price adjustment. However,
Blinder also asked why there was any lag at all. The most popular
answer was a fear of getting out of line with competitors by being the
first to raise prices after costs increased. This would seem to imply a
faster response to cost decreases, since early response in this case
would confer competitive advantage. A survey of New Zealand businesses (Buckle and Carlson 1996) yields the more traditional "prices
rise faster" asymmetry.' So the businesspeople's answers appear to
cover every logical possibility.
Economic theory suggests no pervasive tendency for prices to respond faster to one kind of cost change than to another. In the paradigmatic price theory we teach, input price increases or decreases
move marginal costs and then prices up or down symmetrically and
reversibly. Usually we embellish these comparative statics results with
adjustment cost or search cost stories to motivate lags in response.
But there is no general reason for these costs to induce asymmetric
lags.
Economists have a well-honed skepticism of lay beliefs about how
markets work when those beliefs conflict with our theory. However,
the fragmentary facts do not support that skepticism in this case.
1 The survey used by Buckle and Carlson does not ask directly about speed of
response as Blinder's did. Instead, it asks separately whether prices were raised or
lowered and, among other things, whether costs increased or decreased in a specific
quarter. They find that price and cost increases are paired more frequently in the
same quarter than price and cost decreases.
468
JOURNAL
OF POLITICAL
ECONOMY
Studies involving gasoline (Karrenbrock 1991; Borenstein, Cameron, and Gilbert 1997), various agricultural products (Karrenbrock
1991), and bank deposit rates (Neumark and Sharpe 1992; Jackson
1997) all find that retail prices respond more quickly to input price
increases than to decreases.2 If that finding was shown to be general
and not just limited to a few case studies, it would point to a serious
gap in a fundamental area of economic theory.
My aim here is precisely to generalize, or at least broaden dramatically, the evidence on how prices respond to cost changes. I examine
literally hundreds of markets involving both producer and consumer
goods to see whether there is any central tendency in the speed with
which output prices respond to cost changes. The title summarizes
the main result: the person in the street is right and we are wrong.
This article is unrepentantly descriptive. I try to develop some facts
that our theory of markets will have to subsume rather than test some
specific hypothesis. Accordingly, Sections II, III, and IV describe the
data I use, how I use them, and what they tell us about the central
question of how fast prices respond to costs.
I distinguish producer from consumer markets in part because
most of the case studies have looked for asymmetries in consumer
markets and because that is where most lay opinion undoubtedly
expects to find them. However, I find about as much asymmetry in
producer goods as in consumer goods.
I also examine the response of one decision maker-a large supermarket chain in Chicago-to
its wholesale costs. In this case, I find
no systematic asymmetry.
Section V then examines two potential problems with the central
results: that they may be due to inflation-induced measurement error and that they may conflict with long-run profit equalization. The
results survive both possibilities.
Section VI fishes for some correlates of asymmetry. For example,
I examine the role of traditional proxies for "market power" (concentration, numbers). The notion that some weakness of competition underlies price asymmetries is commonly mentioned in the case
studies (Karrenbrock 1991) and sometimes finds support (Neumark
and Sharpe 1992; Jackson 1997). However, I find that attributing
asymmetries to imperfect competition is unlikely to be rewarding.
There are other negative findings: for example, on the roles of inventories and inflation-related asymmetric "menu costs" of price
changes. And there are some positive findings: on input price volatility and the structure of intermediary markets.
2
For deposit rates the finding is that they respond more promptly to falling money
market rates than to rising market rates.
PRICES
RISE
469
Section VII tries to focus the obvious challenge posed by my results
to our theory of markets.
II.
Data
I analyze three samples of price data. Each contains numerous time
series of output prices matched to input prices. Two of the samples
come from publicly available Bureau of Labor Statistics (BLS) data;
the third is taken from a University of Chicago Graduate School of
Business database supplied by a local supermarket chain.
I use all the data to answer questions such as, If costs go up today,
how long does it take for prices to go up and by how much do prices
rise? I use coefficients from regressions explaining price changes
with current and past cost changes to answer this question. To implement this I have to link some price measure with a cost measure
within a relevant market. Case studies usually do this with coinparatively rich data on one product. For example, Borenstein et al.
(1997) analyze weekly, city-level retail and wholesale gasoline prices.
This level of disaggregation is appropriate if the price adjustment is
completed in a few weeks and markets are local.
My goal of examining many markets at once requires a different
approach. It would be impracticable to investigate the appropriate
time-space aggregation for each of hundreds of products let alone
hope to find corresponding data. So I use the cruder data-monthly
national averages-that are available for many goods and focus on
the central tendencies emerging from analysis of the sample rather
than on results for specific goods. I presume that any errors arising
from the crudeness of the data are not systematic and become small
by averaging.
Here I describe the main features of the three samples, leaving
details to the Appendix.
A.
The ProducerPrice Sample
The BLS producer price index (PPI) is an aggregate of over 1,500
components. Each component is a monthly index of the national
average price for some producer good. The price pertains to the
first transaction after production of the good. This is a transaction
between firms rather than between businesses and consumers. I use
the fact that some of these producer goods are inputs in the production of others to study the transmission of cost changes within the
producer goods sector.
One way of doing this would be to aggregate some appropriate
PPIs (e.g., PPIs for flour, sugar, and energy) into a single input cost
470
JOURNAL
OF POLITICAL
ECONOMY
index to "explain" another (the bread PPI). However, for tractability (and comparability with the skimpy case study literature), I do
something simpler: I analyze only those outputs in which a single
input accounts for a significant fraction (over .2) of the output's
value. Then I use the PPI for that input to explain the output PPI.
One cost of this simplicity is an unrepresentative sample. It tends to
be skewed toward outputs produced with simple technologies in
which the input is heated, crushed, bent, pummeled, butchered, and
so forth. By construction, the sample excludes high-value-added
products.
The Appendix gives a detailed description of how the sample was
put together and how I resolved data problems. Briefly, I began with
the input-output "use" tables, which give each (approximate threeor four-digit) industry's purchases from every other. Pairs in which
the cost share (industry i's purchases from i/i's shipments) exceeded .2 were pursued further using the more detailed data in the
Census of Manufactures. I used the census to refine the matches and
the cost shares to the least aggregated level permitted by the data.
In some cases I consulted industry sources to identify appropriate
matches. Finally, I used a linkage between the standard industrial
classification (SIC) and PPI commodity codes provided to me by the
BLS to identify the specific PPI indexes to be matched. This process
resulted in a sample of 165 input-output pairs with sufficient data
for my purposes.
Generally the sample period is 1978-96. The starting point coincides with the BLS's reform of the PPI to more accurately measure
transaction prices. While double-digit inflation characterizes the
early part of this period, around 85 percent of the sample period is
from the subsequent mild inflation regime.3 (Section V reports results from only the latter period.)
Table 1 provides an industrial distribution of the inputs and outputs and gives some relevant summary statistics. The concentration
in food and metal products (over half the outputs) is evident. However, most of "low-tech" manufacturing (roughly, two-digit SICs less
than 35) is represented. There is enough price volatility in this sample to permit extracting any asymmetries from the data: input and
output prices change in most months, and nontrivial changes in
both directions are common.
B.
ConsumerPrice Sample
Like the PPI, the BLS consumer price index (CPI) is an aggregate
of numerous component indexes. Each measures the retail price of
'The double-digit PPI inflation broke suddenly in the spring of 1981. Subsequently, PPI inflation has averaged under 2 percent per year.
PRICES RISE
471
TABLE 1
SUMMARY
OF PRODUCER
PRICE SAMPLE:
A. INDUSTRIAL
165 INPUT-OUTPUT
DISTRIBUTION
PERCENTAGE
INDUSTRY
Food, agriculture
Textiles, leather
Crude oil
Wood, lumber
Paper
Chemicals
Petroleum, rubber
Stone
Steel
Nonferrous metals
Total
PAIRS
Two-DIGIT SICS
INCLUDED
OF SAMPLE
Outputs
20, 01, 02
22, 23, 31
13
24
26
28
29, 30
32
33, 34
33, 34
Inputs
21.8
14.5
28.5
5.5
.
3.6
6.7
9.1
8.5
6.1
.6
17.0
15.8
100
6.7
9.1
7.3
6.1
.6
17.0
15.8
100
Quartile 1
Quartile 3
B. SAMPLE CHARACTERISTICS
Mean
1. Frequency of Price Changes
(Percentage of All Months)
Outputs:
Increases
Decreases
Inputs:
Increases
Decreases
43.8
34.1
37.2
23.9
52.0
45.4
47.2
40.1
41.8
34.5
52.5
47.1
2. Standard Deviation of Monthly Price Changes
(Log X 100)
Outputs
Inputs
2.7
4.1
1.0
1.5
3. Input Cost/Output
.429
4.2
5.9
Shipments
.31
.50
a specific good or service. Since each of the goods in the CPI has
in principle some counterpart in the PPI, I constructed a sample in
which these counterparts were matched. The goal in doing this is
to estimate how retail markets convert cost shocks specific to themchanges in the producer prices of finished goods-into
price
changes faced by consumers.4 As discussed more fully in the Appen'I did not pursue the chain of production further to ask how primary product
price changes (e.g., crude oil, cattle) ultimately feed into consumer goods prices
(retail gasoline, meat). This would have been feasible for only a handful of consumer
products in which the primary product is a sufficiently important cost component.
Nor did I pursue the chain of distribution. Systematic data on prices paid by retailers
472
JOURNAL
OF POLITICAL
ECONOMY
dix, conceptual differences in the construction of the CPI and PPI
preclude a direct matching; there is no "retail widgets" price index
that is the precise counterpart of a "manufactured widgets" price
index. Substantively, the least aggregated CPI indexes tend to be
more aggregated than the corresponding PPI indexes. This makes
the consumer price sample smaller than the producer price sample.
Also, I usually had to aggregate PPIs to obtain a match with the corresponding CPI index. Neither the items nor the weights (factory shipments) used in the aggregation necessarily correspond to those in
the CPI index. And, I dropped some CPI indexes because the degree
of aggregation seemed too great for the purpose at hand.5
I matched 77 CPI indexes with PPI counterparts. Table 2 gives
some salient characteristics of this sample. It is heavily weighted toward food items, mainly because they are numerically overrepresented among the detailed CPI indexes. Nevertheless, the sample
items account for around half of total consumer spending on physical goods.6 As in the producer price sample, there is no shortage of
"price action" in this sample. For a typical item, both producer and
consumer price indexes change about nine months in every 10, and
there is considerable volatility. At this level of disaggregation, neither the CPIs nor PPIs look anything like the smoothly rising aggregate series familiar to most of us.
C. SupermarketPrices
I have data on the retail and wholesale prices of individual items at
the universal product code (UPC) level sold by the second-largest
supermarket chain in the Chicago area. The items come from a selection of packaged good (i.e., not fresh produce, meat, etc.) categories accounting for about a third of total store sales, and the wholesale price is an average of recent transaction prices rather than the
last transaction prices. Retail prices for a UPC can vary across the
stores in the chain. For most UPC/store pairs I have retail and
wholesale prices for 65 "months" (four-week periods) from September 1989 through September 1994.
Generally, I tried to select the five largest-selling UPCs from each
category and obtain their prices from a stratified random sample of
four stores. I obtained usable data from 24 product categories (listed
in the Appendix), so the maximum number of store/UPC pairs was
to wholesalers are unavailable. So the PPI is, strictly speaking, a retailer input cost
index only when retailers buy directly from the manufacturer.
'Examples would include sewing materials, notions, and luggage or lawn equipment, power tools, and other hardware.
6
list of the sample items and corresponding PPI indexes is available on request.
473
PRICES RISE
TABLE 2
SUMMARY
OF CONSUMER
A. COMMODITY
PRICE SAMPLE
CATEGORIES
PERCENTAGE OF ALL
CONSUMER SPENDING ON
FOR BY
GOODS ACCOUNTED
ITEMS IN:
NUMBER OF ITEMS
IN SAMPLE
Category
Sample
43
13.9
35.6
4
3.5
3.5
Fuel
Apparel
4
6
7.0
3.9
12.6
Recreation
nondurables
Other
4
2.6
4
Automotive
Household
3
9
2.1
12.9
3.1
77
49.0
Food
Alcoholic
beverages
durables
Total
B.
Quartile
1. Frequency
(Percentage
1
Quartile
3
of Price Changes
of All Months)
58.5
37.5
51.4
28.4
66.4
46.8
54.0
33.7
47.8
21.1
61.6
46.6
prices:
Increases
Decreases
2. Standard
Consumer
Producer
100
prices:
Increases
Decreases
Producer
4.7
10.0
16.6
9.7
SAMPLE CHARACTERISTICS
Mean
Consumer
7.2
prices
prices
2.0
3.7
Deviation
of Monthly
X 100)
(Log
.7
.8
Price
Changes
2.1
3.9
480 (five UPCs X four stores X 24 categories). Of these, 357 appear
in my sample. One reason for dropping a UPC was paucity of wholesale price changes, so the sample is biased toward items with frequent price changes.
I formed two subsamples from these data. One uses the store/
UPC pair as the primary unit of analysis. Here I answer the question,
How does the price of a specific item-an 18-ounce box of Kellogg's
Corn Flakes-at a specific location respond to a change in that
item's wholesale price? Because price changes across the chain's
stores (and, to a lesser extent, price changes across UPCs) are not
independent, there are fewer than 357 "real" degrees of freedom
474
JOURNAL
OF POLITICAL
ECONOMY
available to answer such questions. Accordingly, subsequent results
for this UPC/store sample use as the unit of analysis product category averages of the results for specific UPC/store pairs.
The second sampling of these data also uses the product category
as the unit of analysis. However, here the question is, How does the
average price of cereal in the Chicago area respond to a change
in average wholesale prices? For this purpose, I compute a single
wholesale and retail price index for each product category. This is
done by first computing price indexes across all UPCs and then averaging them across all UPCs within a product category (so the category index gives equal weight to each UPC).
The "category average" sample excludes categories with only a
single UPC and categories in which the averaging process leaves too
few (non-de minimus) wholesale price changes for my purposes. Sixteen categories remain in this sample.
Table 3 summarizes some salient characteristics of the two samples. The overwhelming impression is one of considerable price volatility. Standard deviations at the UPC level are two to four times
larger than the CPI or PPI indexes in tables 1 and 2. The averaging
across the UPCs of course reduces price volatility in the category
average sample, but it remains larger than the CPI or PPI indexes.
III.
Empirical Procedure
For each input-output pair in each of the three price samples, I fit
two time-series models to logs of the price indexes. The first is a
simple distributed lag of the general form
(change in output price)t
K
bi i (change in input price),-i
=-,
l
i=O K
+
Ct_i-
(D. change in input price)t i
i=O
+ other variables,
where D = + 1 if the change in input price is greater than zero and
equals zero otherwise. Here current and lagged changes in the price
of some input are allowed to affect an output price asymmetrically
and with a lag. Since (1) is, in principle, a reduced form, it allows
for "other (supply and demand shifter) variables" in addition to
the supply-side shock captured by the input price change.
for
4.
b. a.
b.
3. 2. 1.
a.
figure.this
ii. i.
ii. i.
category.
wholesale
The
At
At
At At
NOTE.-In
Retail
The
Number
sample
Standard
pricesthe
changes
Percentage
UPC/store
of
Wholesale
forare
in retailDecreasing
retailIncreasing
category
sample
of
CHARACTERISTIC
each
based
average
on UPC/store
excludes
store/UPC
sample
sample,
averages
is
pair the
categories
in
unit
based
a within
with
of
on
only
one product
one category.
wholesale
wholesale
pairs
log
deviation
prices: prices:
percategories
months
prices
X
with:
(monthly
category
100):
observation
UPC.Then
8.37.5
is
I
wholesale
a
categories.
34.3
30.9 36.7
34.0
24
14.9 Mean
and For
single
average
one
them item
to example,
retail
to (UPC)
price
obtainat
a
a obtain
5.84.9
30.5
26.3
24.0 32.0
CHARACTERISTICS
12
1
UPC/Store
index
thesingle
for
category
34.3
store.
each
average
percent
There
TABLE
3
9.88.8
40.8
38.0
36.0 41.0
SUPERMARKET
20
Quartile
3
category.are
figure.
figure
up
in to
These
Finally,
five
I row
3ai,
indexes
I items
average
first
and
the
aggregate
24 four
obtain
OF
Quartile
PRICE
SAMPLE
SAMPLES
4.44.5
42.5
42.4
40.6 44.9
3.12.7
39.5
41.0
35.0 41.3
16
Mean
thestores
prices
category
per
for
all
percentage
theaverages
product
to of
obtain
category.
months
store/UPC
the
The
with
pairs
34.3
data 6.15.8
within
shown
a percent
increasing
Category
Quartile
1
Average
46.5
44.0
46.7 48.0
Quartile
3
476
JOURNAL
OF POLITICAL
ECONOMY
The distributed lag model has the virtue of straightforward simplicity. However, my sketchy treatment of the "other variables" (see
below) may mean that the distributed lag underutilizes the available
price data. For example, suppose that output price responds gradually to unmeasured demand or supply shocks as well as to the specific
input price changes in the regression. This would show up in an
autoregressive process in the output price changes. Therefore, to
allow for a more flexible adjustment process, I also estimate vector
autoregression (VAR) models of the general form
(change in output price)t
=
same terms as in (1)
N
+
(2)
et-i (change in output price)t-i
i=l
+ fJ (level of output price
-
equilibrium level) t-i.
The additional terms in (2) allow for some mixture of an autoregressive process and an "error-correction" process. The autoregressive process is represented by the lagged output price change terms.
The coefficients of these terms cannot be signed a priori.
The error-correction term (the last term in [2]) also allows for
entry or exit. If prices have drifted away from equilibrium values
because of incomplete adaptation to past shocks, then current supply changes should move them back toward equilibrium. Accordingly, I expectfto be negative. Operationally, I set the error-correction term, with one exception, equal to the lagged value of (log
output price index - log input price index). This presumes a fixed
equilibrium markup of output prices over input prices for each
good.7
The time-series analysis seeks to answer a general question: How
frequently do prices respond asymmetrically to cost shocks? Accordingly, I eschew any attempt to custom-tailor these models to the circumstances of individual markets. Instead I fit the same modelthe same lag structure and the same list of other variables-to each
input-output pair within any of the three price samples.
7In the UPC sample of supermarket prices, however, I use the log of an index
of average retail price of other UPCs in the same product category, rather than the
UPC's wholesale price, as the proxy for equilibrium price. In preliminary work I
found that for an individual item, such as Kellogg's Corn Flakes, the fact that its
price was out of line with the price of other cereals was more germane than its own
markup.
PRICES
RISE
477
This common model is determined pragmatically: it allows as
many lags as are needed to more or less fully describe the price adjustment process within any sample.8 For example, subsequent results from the distributed lag model for the 77 CPI indexes come
from 77 estimates of (1) each with K = 4. I chose K = 4 because
the number of significant coefficients obtained by adding more lag
terms to the 77 regressions was only about what one would expect
by chance.9 Undoubtedly, K = 4 is too big for some markets and too
small for others. But this is not systematically concealing information
and, for my purpose, seems preferable to conducting 77 specification searches.
For the VAR model, a similar procedure led me to set K = 5 and
N = 5 (i.e., estimate [2] with six current and lagged input price
changes and five autoregressive terms) for each of the 77 CPI regressions. I repeated this algorithm on the PPI and supermarket price
samples to obtain the K and N values appropriate for each of these
samples.
A similar pragmatism drove my choice of the "other variables" in
(1) and (2). They should include other cost shocks plus demand
shifters. It was impracticable to customize a list of such variables for
each of the hundreds of markets analyzed. So I looked only for some
broad aggregate measures that could be added to all the regressions
in a group. That search led to inclusion of the current and three
lagged changes in the log of (1) the PPI for all finished goods less
food and energy, (2) the industrial production index in all the PPI
regressions, and (3) the CPI for all items less food and energy in
all the CPI regressions. These variables summarize the impact of
economywide nominal demand or cost changes, and a sufficient
number of the coefficients were significant to warrant inclusion in
the regressions.10 However, no essential subsequent result would
change if any or all of them were dropped.
8 I did this by estimating
preliminary sets of regressions without asymmetries. Each
set imposed the same lag length (the Kor Nin eqq. [1] and [2]) on all the price
pairs in a group beginning with K = 0 or N = 1. I then increased the value of K
or N progressively until marginal explanatory power seemed exhausted. The resulting value of K or N was then imposed on all estimates of (1) or (2) for that
sample. For the VAR model, I determined N and K iteratively. I began with the K
that worked for the distributed lag and then varied N as described. Once N was
determined, I allowed K to vary further, etc.
9 If 77 regression coefficients are drawn from a random process with a zero mean,
there should be around 10 with Iti ratios greater than 1.5, four with Iti greater than
2.0, etc.
10 I used the "less food and energy" versions of the aggregate price indexes to
reduce potential collinearity and double counting. Food and energy items are prominent among the inputs and outputs in both the CPI and PPI samples.
478
JOURNAL
OF POLITICAL
ECONOMY
Finally, each regression includes month dummies to deseasonalize
the input and output prices."
I use the regression results to answer two questions: (1) How common is asymmetric price response to costs? (2) How large is any such
asymmetry?
For the distributed lag model (eq. [1]), the answers come directly
from the coefficients on the input price changes. In that model, the
cumulative response after k months to an input price decrease in
month t is
k
bt i,
(3)
i=O
and the cumulative response to a price increase is
k
(bti + t-i)
(4)
i=O
So if the difference between these responses,
E
ct-i,
(5)
is positive, there is a positive asymmetry: output prices respond more
fully to a positive cost shock over the k-month period. And (5) gives
the magnitude of the asymmetry. For the VAR model (eq. [2]), it
is necessary to take account of the feedback between output price
changes today and tomorrow (via the autoregressive and error-correction terms). Accordingly, for each input-output pair in each sample, I used the VAR regression coefficients to estimate cumulative
responses to +1 and -1 input price shocks, respectively.12 Then I
estimated the asymmetry after k months as
k
k
AP+ t=O
A
(6)
t=O
where AP is the estimated response of the output price in month t
+ i to a +1 or -1 change in the input price in month t.
" These dummies do not appear in the supermarket price regressions. Preliminary work showed essentially no seasonal price patterns in that sample; it also has
the fewest degrees of freedom per regression.
12 However, I constrained the error-correction coefficient to zero if its estimated
value was positive. A positive coefficient implies implausibly that prices do not converge to equilibrium. Empirically, imposing the constraint makes no substantive difference: (1) positive error-correction coefficients occurred in less than 10 percent
of the regressions in any price sample; (2) when I removed the constraint, none of
the main results derived from the VAR estimates changed; and (3) as elaborated
later, these results are, in any case, essentially the same whether the distributed lag
or VAR is used.
479
PRICES RISE
In practice, (6) and (5) yielded essentially identical results. Thus
only the results from (6) are shown in the next section.
IV.
A.
Estimates of Asymmetric Price Response
ProducerPrice Sample
Table 4 summarizes estimates of equation (6) for 165 producer
goods in which a single input accounts for at least 20 percent of the
TABLE 4
ASYMMETRIES
IN RESPONSE
OF PRODUCER
PRICES TO INPUT
CUMULATIVE
COST CHANGES
RESPONSE
AFTER MONTH:
SAMPLE
SIZE
0
2
4
8
A. Mean Response to Input Cost Changes
(Full Sample)
Input cost change = + 1
Input cost change = -1
165
165
.235
.127
.371
.233
.430
.270
.505
.354
B. Asymmetries in Response to Unit Cost Changes
1. Mean Asymmetry
Full samplea
165
Input cost share > .4
80
Input cost share ' .4
85
.108
(5.0)
.147
(4.8)
.071
(2.4)
.139
(5.0)
.129
(3.2)
.148
(3.9)
.160
(4.6)
.130
(2.4)
.189
(4.4)
.150
(3.4)
.093
(1.6)
.205
(3.0)
2. Share of Sample with Asymmetry Greater than Zero
Full sample
165
Input cost share > .4
80
Input cost share ' .4
85
.69
(5.1)
.69
(3.6)
.68
(3.6)
.72
(6.1)
.76
(5.5)
.67
(3.3)
.72
(6.3)
.70
(3.9)
.74
(.51)
.58
(2.1)
.60
(1.8)
.59
(1.7)
3. Mean Asymmetries by Industry of Output
Food
Textiles/leather
Woodb
Paper
Chemicals
Petroleum/rubber
Nonferrous metals
Steel
36
24
12
15
14
10
26
28
.168*
.087
.222*
-.124
.129
.008
.060
.196*
.086
.042
.185
.273*
.263*
.078
.106
.187*
.087
.032
.124
.413*
.173
.111
.115
.297*
.142
.010
.176
.434*
.236
-.099
-.027
.331*
NOTE.-Data come from the VAR model. Results for the distributed lag model mimic those shown here.
In panel A, the t-ratios for all mean responses exceed 6.6. In part 2 of panel B, the t-ratios (in parentheses)
pertain to the difference from .5.
* Itl > 2.0.
aRow I minus row 2 in panel A.
b Includes one stone, clay, and glass item.
480
JOURNAL
OF POLITICAL
ECONOMY
good's value. Panel A gives estimates of the average values of each
of the two terms in (6), and panel B focuses on the difference between the two. For example, the first row of panel A says that a 1
percent input price increase in t = 0 leads to a 0.235 percent output
price increase in the same month for the average good in the sample. After eight more months, the price of this good has risen a total
of 0.505 percent.13
The results in panel B point to a single conclusion: positive asymmetry is a fact of life in industrial markets. It is pervasive: over twothirds of the sample markets exhibit positive asymmetry in t = 0. It
is large: the t = 0 response is nearly twice as great when input prices
rise as when they fall. And the phenomenon lingers: only at t = 8
is there even faint evidence of a narrowing in the gap between responses to input price increases and decreases. That gap ought to
disappear entirely, else input price volatility would produce rising
price-cost margins over time. But it is not doing so within the period
in which I can find measurable responses to the cost shock. Section
VI attempts to resolve this puzzle.
The preceding results all hold when the sample is carved up by
input cost share'4 and by industry. In the latter case, the individual
estimates in section 3 of panel B are often insignificant because of
the small size of the subsamples. But the noteworthy feature here is
the overwhelming number of positive point estimates (29 of 32).
Clearly the phenomenon of positive asymmetry is not confined to
only a few kinds of industrial products.
B.
ConsumerPrice Index Sample
Table 5 repeats the preceding analysis for the sample of 77 consumer
goods for which I can match a CPI index to a PPI "input" index. The
results are remarkable more for their similarity to producer goods
markets than for any differences. Once again an overwhelming majority of these consumer markets exhibit positive asymmetry. The
size, persistence, and pervasiveness of the positive asymmetries here
essentially mirror those in producer goods markets. If anything,
these characteristics are a bit stronger for consumer goods. But no
relevant test would reject the equality of the asymmetries in con-
13 To put the number in perspective, note that the average input cost share in
this sample is .43, which would be the expected average long-run response under,
say, constant cost Cobb-Douglas or Leontief production. The average of the positive
and negative price responses (.505 and .354, respectively) roughly equals the average
input cost share.
4 None of the differences between rows b and c in panel B are significant.
481
PRICES RISE
TABLE 5
ASYMMETRIES
IN RESPONSE
OF CONSUMER PRICES
PRICES
TO CHANGES
IN PRODUCER
AFTER
CUMULATIVE RESPONSE
MONTH:
SAMPLE
0
SIZE
A. Mean
Producer
Producer
price change
price change
=
=
+1
-1
77
77
.194
.067
77
Share of sample with asymmetry greater than zero
Mean
asymmetries
consumer
by type
good:
77
.368
.159
.127
in
Prices
.482
.274
in Response
to Unit
Producer
Prices
B. Asymmetries
Mean asymmetrya
to Changes
Response
Producer
5
3
1
.209
(5.4)
(5.3)
.766
(5.5)
(7.7)
.831
.522
.336
in
Changes
.209
.186
(4.4)
(4.0)
.740
(4.8)
.662
(3.0)
of
Food:
Fresh
Processed
Fuel
Furniture/appliances
Apparel/jewelry
Automotive
Drugs/
toiletries
Recreation
17
30
4
7
.016
.144*
.205*
.275*
.086
.289*
.096
.355*
.097
.276*
.069
.382*
.106
.251*
- .036
.446*
8
.281*
.285*
.319*
.339*
3
4
4
.066
- .036
.031
.167
- .013
.095
.141
.055
.000
.069
- .309
.081
NOTE.-t-ratios are in parentheses.
* Itl > 2.0.
'Row I minus row 2 in panel A.
sumer and producer markets at similar stages of the adjustment process.
C. SupermarketPrices
There is a sharp contrast between the preceding results and those
for prices at a specific supermarket chain as summarized in table 6.
Here the unit of observation is the product category, and the table
gives summary statistics across the sample categories. These statistics
are shown for two distinct levels of aggregation. In the UPC/store
sample, equations (1) and (2) were estimated for 357 pairs of individual UPCs at specific stores. Then I used the regression coefficients
to generate estimates of the cumulative response to wholesale price
changes for each UPC/store pair. Finally, I average these estimates
482
JOURNAL OF POLITICAL ECONOMY
TABLE 6
ASYMMETRIES
IN RESPONSE
OF PRICES AT A SUPERMARKET
ITS WHOLESALE PRICES
CHAIN
TO CHANGES
CUMULATIVE
RESPONSE
MONTH:
IN
AFTER
SAMPLE
SIZE
A. Mean
UPC
0
Response
1
to Wholesale
price change
=
+1
24
.593
(3.9)
Wholesale
price change
=
-1
24
.585
(6.0)
Product
average
sample:
Wholesale
price change
=
+1
16
.445
(4.2)
Wholesale
price change
=
-1
16
.447
(7.6)
B. Asymmetries
Cost
Changes
.415
(4.7)
.460
(3.8)
.418
(6.7)
.406
(5.0)
in Response
Changes
.565
(4.8)
.394
(5.1)
.473
(6.8)
.461
(6.1)
.405
(6.1)
.423
(4.9)
.447
(6.7)
.508
(7.1)
to Unit
asymmetry:
UPC samplea
Product
average
24
sample
16
.011
(.1)
-.004
(.0)
Share
Price
sample:
Wholesale
Mean
3
2
-.045
(.7)
.016
(.2)
.170
(1.8)
-.017
(.3)
.012
-.061
(.1)
(.7)
of sample
with asymmetry
than zero:
greater
UPC sample
Product
average
24
sample
16
.42
.42
.58
.54
(.8)
(.8)
(.8)
(.4)
.41
.53
.41
.65
(.7)
(.2)
(.7)
(1.3)
NOTE.-t-ratios are in parentheses.
a Row I minus row 2 in panel A.
over all UPC/store pairs in a product category. That average answers
a question such as, How quickly does a change in the wholesale price
of a typical brand of cereal get reflected in the retail price of that
brand at a specific store?
For the "category average" sample, only one regression is run for
each product category. It uses indexes of average wholesale and retail prices across all the UPC/store pairs in that category. So it answers the question, How does the average price of cereal at this chain
respond to changes in the chain's average wholesale price of cereal?
However, as table 6 shows, the degree of aggregation makes no
difference in practice for the behavior of interest here. The most
prominent feature of the table is the absence of any systematic asymmetry in panel B. The plethora of insignificant differences from symmetry speaks with one voice in this respect.
PRICES RISE
483
The results in table 6 describe only one decision maker. However,
if they were general, the implication would be that asymmetry at the
market level is not produced by some simple aggregation of individual decisions. Instead these results would suggest that some subtlety
of the interaction of these individual decisions produces the asymmetry. This would, of course, hardly be the first time in economics
that the distinction between individual and market behavior was important.
V.
Potential Problems
A.
MeasurementError
All the input prices used in the preceding analyses are undoubtedly
measured with error. Most of them also drift upward. This combination of measurement error and secular inflation could produce spurious positive asymmetries by making measured price decreases less
accurate than measured increases.15
However, inflation can also produce the opposite bias. The main
issue is whether inflation systematically leads to a lower signal-tonoise ratio (variance of true price changes/variance of measurement error) for measured input price decreases, and inflation can
move the relevant relative variance ratios in either direction.16 Then
it is easy to conjure cases in which inflation would introduce a negative bias into estimates of asymmetry.'7
'5An extreme example would be a "true" input price that sometimes rose but
never fell. If we measured that price with some random error, then everymeasured
price decrease would be pure noise. In this case, my method would produce an
estimated asymmetry even if there were none. The expected coefficient on price
decreases would be zero because the price decreases are entirely noise.
16 In general, if the mean change is positive, the variance of measured input price
increases will exceed the variance of decreases. This leaves room for the variance
of boththe true and error components to be larger for increases than for decreases.
For example, suppose that the true changes are uniformly distributed with a mean
of + 1 over [-1, 3] but they are measured with an error uniformly distributed over
[-2, 2]. Thus the measured changes are distributed over the range [-3, 5]. The
larger range of measured increases (here 5 vs. 3) can therefore accommodate the
whole range of measurement error. However, the error consistent with a measured
price decrease has an upper bound of only +1.
17Consider the previous example: the mean true change equals +1 uniformly
distributed over the range [-1, 3], but the error is distributed uniformly over the
range [-3, 3]. Then measured input price changes range between -4 and 6, and
we have the following ranges for the true and error components: For measured
increases in the range [0, 6], the range for the true components is [-1, 3] and for
the error components is [-3, 3]. For measured decreases in the range [-4, 0], the
range for the true components is [-1, 3] and for the error components is [-3, 1].
Here the largest error in measured decreases is + 1. Consequently, the ratio of the
range of true to error components (here equal to the ratio of standard deviations)
is 4/4 = 1 for measured decreases and only 4/6 = .67 for measured increases. In
this case, using the measured input price changes to explain output price changes
484
JOURNAL OF POLITICAL ECONOMY
The lack of an a priori bias from inflation does not, of course,
rule out an actual bias. However, the underlying data do not suggest
that such a bias is driving the results. If such a bias is important on
average, it should be more important the greater the drift of true
nominal input prices is (and consequently the more abnormal nominal price decreases are). There is no such pattern in my data when
I use the average monthly change over 18 years as an estimate of
the true drift of nominal input prices. All the simple correlations
between the drift of input prices and any asymmetry measure are
indistinguishable from zero in both samples. Two nonparametric
tests of this relation yielded the same negative finding (as more complete tests in the next section do)."8
Finally, I reestimated all the asymmetries on data exclusively from
the low-inflation period, 1982-96. They should be systematically
lower if inflation-induced measurement error is driving the results.
However, as shown in table 7, the results for 1982-96 are indistinguishable from the full-period results.
B.
Long-Run Equilibrium
I find no evidence that asymmetries eventually disappear and only
weak hints that they even get smaller over time. Taken at face value,
such results imply permanent effects on price-cost margins from
nominal cost shocks. For example, a series of offsetting cost shocks
would produce cumulatively higher margins. This implication conflicts so much with basic equilibrium concepts as to raise questions
about the results. For example, could the results reflect some other
force positively correlated with long-period margin changes?
Unfortunately, theory is silent on exactly when the long run arrives. So the most straightforward answer to such questions would
be that the asymmetries do eventually disappear, but do so too slowly
would result in moredownward bias in the coefficient for price increases, i.e., a negative asymmetry.
18
For the first test I carved the distribution of input price changes into quartiles.
Then I regressed each asymmetry measure in both the producer and consumer
goods samples on dummies for the second through fourth quartiles. If there is an
inflation bias, then the coefficients of these dummies should be positive and increasing as we move from lower to higher quartiles, and the intercept (the first quartile
asymmetry) should be consistently smaller than the average asymmetry. No such
pattern appears in the data. For example, none of the 24 coefficients of the dummies
were significant, and 11 were negative. None of the intercepts were distinguishable
from the average asymmetry. In the second test, I regressed each of the estimated
asymmetries on the inverse rank of the input's inflation rate. Here the symptoms
of inflation-induced measurement error would be a significantly positive coefficient
on the rank variable and a constant below the mean asymmetry. Neither turned up
in any of the eight regressions.
PRICES
485
RISE
TABLE 7
ASYMMETRIES
IN PRODUCER
AND CONSUMER
INFLATION
PRICES:
A. PRODUCER
RESPONSE
8
.139
.160
.150
.135
(4.8)
.175
(4.7)
.171
(3.3)
2
Mean
period
(table
4)
.108
.100
(4.7)
onlya
Share
Full
period
(table
1982-96
onlya
4)
.69
.68
B. CONSUMER
of Sample
with Asymmetry
Greater
than Zero
0
.72
.70
(5.6)
RESPONSE
.127
.143
.209
.230
(5.4)
(5.5)
Share
Full
period
(table
1982-96
only
5)
.77
.73
(4.5)
(1.8)
AFTER MONTH:
1
Mean
5)
.58
.57
PRICES
CUMULATIVE
period
(table
1982-96
only
Asymmetry
.72
.69
(5.2)
(4.8)
Full
AFTER MONTH:
4
0
1982-96
FROM THE Low-
PRICES
CUMULATIVE
Full
ESTIMATES
1982-96
PERIOD,
3
5
Asymmetry
.209
.229
(4.5)
.186
.222
(4.2)
of Sample
with Asymmetry
Greater
than Zero
.83
.81
(6.8)
.74
.73
(4.5)
.66
.68
(3.3)
NOTE.-t-ratios are in parentheses.
'Sample size equals 164 vs. 165 in table 4. One item had too few observationis to estimate the regression.
to be picked up in my data. This answer seems consistent with the
data.
Specifically, I focus on long-period trends in output and input
prices since they should be dominated by long-run equilibrium conditions. If the asymmetries were not ultimately reversed, output and
input prices would tend to drift apart for a typical good. Moreover,
the larger the asymmetry, the greater the divergence between output
and input price trends.
The crude data are, however, inconsistent with these implications.
For example, when post-1981 data are used, the price of the average
producer good in my sample rises 16 basis points per month (bppm)
486
JOURNAL
OF POLITICAL
ECONOMY
whereas input prices drift upward at a statistically indistinguishable
14 bppm. For consumer goods, the comparable figures are 22 and
13 bppm, and this difference is significant.
Since retail-wholesale margins can change for many reasons, the
more germane question is whether the difference between output
and input price trends is positively correlated with the degree of
asymmetry. In fact the correlations are insignificantly negative: -.14
for producer goods and -.19 for consumer goods (when the eightand five-month asymmetries, respectively, are used). This lack of
positive correlation suggests that the positive asymmetries are not
permanently ratcheting up output prices relative to input prices.
These correlations are only suggestive because my data account
for the price of only one input. To provide a sharper test for the
long-run effect of asymmetries, I decompose the long-run trend of
the nominal price of an output (y) as
y = (1-c)k
+ cx + e,
(7)
where x is the long-run trend in the price of the single input whose
price I measure, c is the cost share of x in production of y, k is the
long-run trend in prices of other inputs into y, and e is other factors,
such as the difference in productivity growth in y relative to inputs.
Equation (7) would describe the long-run equilibrium exactly (e
= 0) with, for example, constant returns Cobb-Douglas or Leontief
production; it would approximate, say, constant elasticity of substitution production with a reasonably small elasticity of substitution. So
it seems a useful starting point.
If the long-run equilibrium of y were not affected by a few months
of asymmetric response to input price changes, then something like
(7) would hold for the typical good. A regression of y on 1 - c and
cx would then yield a constant equal to zero, estimated k as the average input inflation rate, and a coefficient of cx of one. Adding the
stimated asymmetry to this regression should yield a coefficient of
zero. However, if y were permanently affected by an asymmetric response to cost shocks, this coefficient would be positive (more asymmetry would create more divergence between input and output
price trends).
Table 8 implements this strategy. It also includes regressions that
interact the asymmetry with the average input price change since
any permanent effect would disappear if positive cost shocks were
pervasive. The important result is that there is no evidence of any
permanent effect of asymmetries on the long-run trend of output
prices: none of the relevant coefficients differ from zero.'9 These
'9 The other implications of long-run equilibrium all hold for producer goods:
the constant is indistinguishable from zero, the estimated k is approximately the
487
PRICES RISE
TABLE 8
ASYMMETRIES
AND LONG-RUN
TRENDS
IN OUTPUT
MEAN
AND INPUT
MONTHLY
PRICES:
1982-96
IN OUTPUT
CHANGE
PRICE (Basis Points)
INDEPENDENT VARIABLE
Constant
1 -
c = 1 -
input cost share
cx = input cost share X input price
trend
Asymmetry (8 months or 5 months)
Producer Goods
Consumer Goods
(N= 164)
(N= 77)
(1)
(2)
(3)
(4)
-.915
(.2)
-.977
(.2)
21.6
(3.4)
21.4
(3.4)
20.5
20.2
(2.8)
.969
(2.8)
.993
(7.8)
(8.1)
-.860
1.40
(.5)
(.6)
Asymmetry X x
Adjusted R2
Standard error of estimate
.30
13.9
-.674
(1.2)
.31
13.9
-13.1
-13.1
(1.0)
1.02
(1.0)
1.02
(7.7)
(7.7)
-1.31
(.4)
.44
11.2
.382
(.1)
-.075
(.4)
.43
11.3
NOTE.-t-ratios are in parentheses. The assymmetry variables are the cumulative asymmetries after T
months summarized in tables 4 and 5 (T = 8 for producer goods; T = 5 for consumer goods). The estimated
cost share for consumer good markets attempts to adjust for the layer of wholesale transactions between
producers and retailers as follows: For each consumer good's product category, census data provide retailers'
cost of goods sold per dollar of sales (rcgs) and the same variable for wholesalers (wcgs). In 1987, economywide wholesale sales were 1.45 times total retail sales (i.e., the average good went through more than one
wholesale transaction). I use this to estimate the input cost share for retailers as rcgs X wcgs' 45.
results imply that the asymmetries do ultimately disappear but that
it takes longer than five or eight months for this to happen.
VI.
Exploratory Analyses of Price Asymmetries
The obvious question raised by the preceding analysis is, Why are
asymmetries so pervasive in real-world markets? Unfortunately, I
have no answer. Instead, I investigate a more limited question: Are
there obvious regularities in the asymmetries? I do this by regressing
the degree of asymmetry on a list of readily available market characteristics in the hope that the larger theoretical question will at least
come into better focus. For example, industrial organization specialists (and noneconomists) would be drawn to theories based on imperfect competition. So I include conventional market structure
and numbers-in
the analysis. Macromeasures-concentration
overall PPI inflation rate, and the coefficient of the cost share-weighted input price
trend is approximately one. Only the latter implication holds for consumer goods.
However, if (7) is simply imposed on consumer goods by constraining the parameters, the key result-no
permanent effect of asymmetry on y-is unaffected.
488
JOURNAL
OF POLITICAL
ECONOMY
economists have focused on "menu-cost" explanations of price
rigidities generally (Sheshinski and Weiss 1993) and inflationinduced asymmetries in menu costs in particular. For example, in
Ball and Mankiw (1994), secular inflation, by reducing real prices,
allows firms to avoid the menu cost of a prompt response to a negative cost shock. Accordingly, I include variables designed to capture
the importance of asymmetric menu costs.
Table 9 describes all the independent variables used in the analysis, and tables 10 and 11 summarize the results. The independent
variables fall into two broad categories: those describing the behavior of input prices and those describing the structure of the inputor output-producing industries. In the case of consumer goods, I
also include some characteristics of the wholesale intermediaries between goods producers and retailers.
Input price behavior variables include the input's cost share, its
price volatility, and two variables related to menu costs: the drift of
input prices and the differential persistence of positive versus negative input cost changes. If input prices are rising and input price
increases tend to be more permanent than decreases, a fast response
to input price decreases could waste menu costs. So menu-cost stories would imply a positive correlation between both variables and
the degree of asymmetry. (Including the drift of input prices also
provides further evidence on the measurement error issues discussed in the preceding section.)
The industry variables include standard measures of competitive
structure (numbers, Herfindahl-Hirschman index [HHI]), market
size, production technology (assets/sales), and inventory holdings.
Inventories proxy for storage costs. They should affect the speed (but
not, in any obvious way, the asymmetry) of response to cost shocks.
Finally, I include measures of the geographic concentration of
suppliers and buyers. The motive here is to see whether propinquity
affects the speed of price propagation. Here, too, there is no reason
to expect any effect to be asymmetric.
Given the theoretical vacuum, I let the data speak by first using
all the available independent variables and then winnowing the list
to those variables that show some "promise" in the first regression.
Only the second of these two steps is shown in tables 10 and 11.
However, the notes to these tables indicate which of the variables
in table 9 did not make the first "cut."20
For producer goods, two positive results emerge from this two20 I chose the lags in tables 10 and 11 to keep the detail manageable and to keep
the information overlap in the regressions reasonably small. Since the asymmetry
measure is cumulative, there is a positive correlation across the various lags. However, they are hardlymirror images. For example, for the three lags in table 10, the
PRICES
RISE
489
stage fishing expedition: (1) less input volatility is associated with
more asymmetry, and (2) the structure of output markets "matters,"
but in a way that resists easy labeling. The negative impact of input
price volatility seems the more robust of the two. It shows up in every
specification in table 10.
Does less competition produce more asymmetry? The answers
from table 10 are conflicting. Two conventional proxies for output
market competition-numbers
and concentration-have
opposing
effects. Fewer competitors are associated with more asymmetry, but
more concentrated markets produce less asymmetry. Since the HHI
is a "numbers equivalent," the difference between the two coefficients is a summary measure of the effects of "more competition." 21
This difference is indistinguishable from zero for all the regressions
in table 10.
If the results do not therefore cry out for imperfect competition
stories, neither do they provide much encouragement for menu-cost
or inventory models. None of the proxies for these latter two are
consistently important.
The negative effect of input price volatility on asymmetry also
shows up consistently in the consumer goods sample in table 11.
This mirrors even in rough magnitude the result for producer goods
markets. These twin results suggest that the mysterious good "delayed response to input price reductions" at least obeys the law of
demand: The "price" of failing to increase supply when input prices
fall is the loss of the extra margin on incremental sales. This price
is greater-and
the results on volatility imply that the output response is apparently less restrained-when
input prices are more
'
volatile.22
Moreover, the effect is empirically important. For example, suppose that price volatility rose by a standard deviation. The regrescorrelation is .22 for months 0 and 2, .26 for months 0 and 8, and .55 for months
2 and 8.
21 With n
equal-sizedfirms, In HHI = -In n. So the difference between the coefficients of In HHI and In nwould give the effect of a 1 percent increase in the number
of equal-sized firms.
22 The discussion here refers to the size of price decreases rather than to overall
volatility.However, empiricallythe two are basicallyindistinguishable. The correlations between the logs of the standard deviation of prices and the mean absolute
price decrease are .95 and .98 for the producer and consumer goods samples,respectively. For example, consider a price taker with the marginal cost function MC =
X, where X is output produced by a single variableinput. If the input price falls by
100 6 percent, the usual storywould have Xgo from P (price) to P7 (1 - 6), and
the profits from the incremental output would be (P6)2/2(1 - 6). If instead the
firm increased output by only a fraction, k, of the difference between P and P7 (1
- 6), it would sacrifice profits of [P6(1 - k)]2/2(1 - 6). For any k, this sacrifice is
increasing in 6, the size of the input price reduction.
CHZ
~~~~~
~~~~~~~c
ct
~~~o~~~~~-b
"e *~~~
P4~~~~~~~~~~~~~~
(LI5
cd4e 4
I
l
~~-
Q
-
Q
bodCQ
~
V~~~~~~~~~~~~~~~~S
~~~~~~~~~~~~.
c
~Q
s-~~~~
5
Vl
L
t
;>
~~~~~~~
~~~~~~~~~~~c
V ~~~~~~~~~~~~V
~z ~~~~~~~~~
H~~~~~~~L -
O
4Vi
4-j
4-i
4"
-
V.
4-i
.5.j
~
r
-
4o J
4"e
4-i
4
~~~~~~~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~
-~~~~~~~~~VC~~~~~~~~~~~
Z
~~~~~~~~~~~~~~~~~~~~~~0~~~~-jc
(L
tJ
v
ctc
_'
~
~~~~~~~~~~(L
--
-
;L4
C
(L)
(L)
V
n
Z
4o
c (V
~~~
ct~
0
490
'
-
-I
V
0
V
t
c
:
-V
b
D S4
t *
of
the
Crude(farm)
Census
Retail
of oil:total industries
Allfarm Trade
Consumer
Four-digit
were SOURCE.-All
Assets/salesMerchandise
inventory
Wholesale
SIC
good data
data
data
Sales/inventories
line
that are
Trade, by estimated
sample
(unpublished
taken or pertain
sales
to
exceptinventory/receipts
data
only.
produces
from came
theHHI
1987,
commodity.
ratio
pertain
from
is
the
and
Sales to
input
Four-digit
Statistical
used
or
or or
1992);
alternate
Abstract
where
PPI:
output
output.
higher
approximate
assets/sales,
of
Stores
Sales
Consumer
Percentage
Geograph
Geographic
Establishments
sales
by
expenditures
(nonmanufacturers)
concentra
concentration:
out-in
manufacturers
(nonmanufacturers)
Sum 1.9 Sum
tion
if
Number Iof mately
in over
Wholesale
of Percentage 5 all
the
of
share
sales
region
all
of
1.
by
percent nine
wholesale
of production
is
output
the
census
wholesale
Minimum
supermarkets
D.
line
by
data)
which
U.S. is
depreciable
line,
products:
sells
fruits
is
unless
with CPI-U
sales
store/sales
unavailable.
industry0 regions
the
sample
(CR)
if
U.S.
Retail establishments
concentration.
and
(1978-96);
vegetable of selling
Iof
whichhandling nonmanufacturers
concentrated
specialized
made
assumed
modal
otherwise
Number
C.
The
index,
Dept.
average.
BLS:
the
to
to
in
same is
that
Census
by
merchandise
of of
fruits
population)
assets/sales
stores
be total
(CW)
Market
share
arethe
the
sold
per
line
sales).
thefirms Bureau
employment
andvegetables
for
of
specified.
employment
by
farm of
- (CM)
not
and
Bureau)
rather
is
least
Wholesale
same
CM:
December
Agriculture,
than
as HHI
including
stores
sectorLabor
merchandise
merchandise
share
than
industry
for
Example:
the
Census
1986(Consumer
are
vegetables
other
of
most
line
line
supermarket)
set
Statistics
allocated
+
populous
Industryt
good.
manufacturers'
at asset/sales
Agricultural
input
X
Good
supermarkets,
(BLS)
apples
types
petroleum
the
manufacturers'
ratio (1988).
heavily divided (sales
sales
are
of
region
exactly
fruit
is Statistics
Manufactures; supermarkets
employme
including
of
by
sample Some
as
industry
Sample)
number
refining.
in
offices
and
CW:
offices
of stores. the
data
(1988),
assumed
or
fruits
their
Scrap
would
for
or
included
maximum Census specialized
region
fruit
be
in
(Mountain,
good.
and
of
to
metal:
and provides
agents
population.
vegetable
applicable
and
the
employment
products
agents
All
I
with
used
Therefore,
to sales,
of
Wholesale
share
inventories.
data
(CW)
Weights
forstores,
each
(CR
(CW)
"fruit
minimum,
of
vegetables
(CM)
=
Trade;
output,
For
come
vegetable
Maximum
etc.and
approximerchandise
and
per
and CR:
is
from product,
apples, (size
unline
populaexamof
Sales
the data
sources
ple,Sales
(CR
fruit
Weight
as
of
of
stores
stores
for
are producer
Weighted
for
store
of
midpoint
United
+
=
and
Acquisition
the are
the and1
published
of
vegetable"
used price
.
the
follows:
stores.
States,
X Supermarkets
the
kind
for
commodity-specific cost
data
apples,
of assumed
index
number
of
most (CR) sell type/size
good
data
except
forsample
vegetable
of
in
fruit
equalare
of
the
Agricultural
business
to HHI, geographic unpublished merchandise
more
merchandise
the
provided
and
period,
heavily
stores
type
good
that the
SIC
which
therespectively.
nonmanufacturing
andsome
Census
JOURNAL
492
TABLE
REGRESSIONS
OF PRICE ASYMMETRIES
ON INPUT
CHARACTERISTICS:
OF POLITICAL
ECONOMY
10
PRICE AND SELECTED
PRODUCER
MARKET
PRICES
LAG
INDEPENDENT
Input
price behavior:
Cost share
In standard
...
deviation
Input
supply
industry
In companies
industry
?
-.084*
...
-.096*
8
+
-.
-.148*
(SIC):
In HHI
Output
2
0
VARIABLE
-
-
.061t
...
...
...
-.035+
-.091*
-.173*
(SIC):
In companies
In shipments
Raw materials/shipments
...
.054t
.113*
Assets/shipments
-.685.
*19
-.l2t
-.189+
In HHI
-.03+
-.051t
-.125*
Geographic
concentration
-.097*
NOTE.-See table 9 for a description of the independent variables. The dependent variable is the estimated
asymmetry over the indicated period. Each regression is weighted by the reciprocal of the standard error of
the first interaction term in eq. (2). Preliminary analysis revealed heteroskedasticity of the residuals from
unweighted regressions, and this was eliminated in the weighted regressions. Equation (2)-the VAR modelis used to estimate dependent variables in the table; the distributed lag model yields essentially identical
results. The two-step method used to select independent variables is described in the text. The following
variables in table 9 did not survive the first step: (1) input price behavior: drift and persistence; (2) input
raw materials/shipments,
supply industry: value of shipments, finished goods/shipments,
assets/shipments,
and geographic concentration; (3) output industry: finished goods/shipments.
Variables with right-skewed
distributions are entered in logs. The regressions include dummies for nonmanufacturing industries in which
independent variables had to be estimated (see table 9).
* Coefficients have jtj > 2.
t Coefficients have 1.5 < ItI < 2.0.
+ Coefficients have 1 < ItI< 1.5.
? Coefficients have Iti < 1.
sions imply that the two-month asymmetry measure would then fall
by about half in both the producer and consumer good samples.
The only other positive result in table 11 is that there appears to be
more asymmetry when the supply chain is more fragmented. When
retailers obtain supplies directly from the manufacturer or from a
few large wholesalers, there is less asymmetry.
These effects are also important empirically as well as statistically.
The coefficients imply that half the mean two-month asymmetry
would be eliminated by a one-standard-deviation change toward less
fragmentation in any one of the three wholesale market variables.
Consumer markets, like producer goods markets, provide little
support for inventory or menu-cost stories.23And none of the retail
23
In fact, the consumer
seem perverse.
For example,
one might
exprice results
for goods
with slim inventories
to behave
more
like a textbook
pect prices
spot
are associated
market.
low manufacturer
inventories
with more
rather
However,
the negative
than less asymmetry.
coefficient
of the (differential)
Similarly,
persisin table 10 is opposite
tence variable
of that implied
of
by a menu-cost
explanation
asymmetry.
493
PRICES RISE
TABLE 11
REGRESSIONS
OF PRICEASYMMETRIES
ON PRODUCERPRICEAND SELECTED
MARKETCHARACTERISTICS:
CONSUMERGOODS
LAG
INDEPENDENT
VARIABLE
0
2
5
Input price behavior:
Cost share
In standard
Persistence
deviation
Producer good industry:
In companies
In HHI
Finished goods/shipments
Wholesale industry:
Percentage sales by manufacturers
In establishments (nonmanufacturers)
In sales (nonmanufacturers)
... ?
- .067*
- .066*
-.137*
-.064*
-.134*
-.061*
-.051*
-.043t
-.948*
-1.328*
-.031+
-.285+
.063+
.063+
-.644*
.206*
...
-.141
*
-.815*
.355*
- .280*
NOTE.-See table 9 for a description of the independent variables. See also the note to table 10. The
regressions are weighted by the reciprocal of the standard error of the first interaction term in eq. (2). The
dependent variable is the estimated asymmetry from the VAR model over the indicated period. The distributed lag model yields essentially identical results. The two-step method used to select independent variables
is described in the text. The following variables in table 9 were eliminated in the first step: (1) producer
price behavior: drift; (2) producer good industry: value of shipments, raw materials/shipments,
assets/shipments, and geographic concentration; (3) wholesale industry: none; (4) retail industry: all. Variables with
right-skewed distributions are entered in log form.
* Coefficients have Iti > 2.
t Coefficients have 1.5 < Iti < 2.0.
Coefficients have 1 < Iti < 1.5.
Coefficients have Iti < 1.
market characteristics-their
size, the density of stores, retailer inventories, and margins-seem to have any connection to the degree
of asymmetry in consumer goods prices.
VII.
Summary and Conclusion
The odds are better than two to one that the price of a good will
react faster to an increase in the price of an important input than
to a decrease. This asymmetry is fairly labeled a "stylized fact." This
fact poses a challenge to theory. The theory of markets is surely a
bedrock of economics. But the evidence in this paper suggests that
the theory is wrong, at least insofar as an asymmetric response to
costs is not its general implication. In this case, lay prejudice comes
closer to the truth than our theory does.
The theoretical puzzle is unlikely to be solved by a roundup of
the usual suspects. Price asymmetry is as characteristic of "competitive" as "oligopoly" market structures. It is found where the buyers
are numerous and unsophisticated consumers as well as where they
are large and presumably sophisticated industrial purchasers. Neither inventory holdings nor menu costs seem a key ingredient in
494
JOURNAL
OF POLITICAL
ECONOMY
producing price asymmetries. The only clear regularity I found was
that more volatile input prices are associated with less price asymmetry.
Perhaps the first path to a solution toward which many economists
will be drawn is "adjustment costs," since we regularly invoke them
to rationalize lags in price response generally. For example, consider
a good produced with inputs (the "materials" studied here, "labor," etc.) all purchased under "at-will" contracts. If the price of
materials rises, inputs will have to be disemployed. This can be done
quickly at low cost given the nature of the contract. However, if the
price of materials falls, new inputs will need to be recruited. And
there are costs to doing this quickly (search, price premia) that are
absent when inputs are disemployed. This asymmetric adjustment
cost story would be consistent with the kind of price asymmetry I
find here.
My findings suggest some caveats and perhaps other paths for future research. I found no asymmetry when I examined the response
of a single decision maker to its own costs. By contrast, I found
above-average asymmetry between factory and consumer prices
when there were many small intermediaries between the factory and
the retailer. This suggests that an explanation for asymmetry may
require a fuller understanding of those vertical market linkages.
My research design focused on the easy cases: those in which a
single input is a major cost component. This made it relatively easy
to equate a "cost shock" to the change in a single, often volatile,
price series. But, at least for producer goods, my procedure results
in a possibly unrepresentative sample of low-tech, low-value-added
items. And it leaves a large question for future research: Do prices
really respond asymmetrically to cost shocks generally? Or is the
asymmetric response limited to one input price in cases in which
the input happens to be important?24
Appendix
The Three Price Samples
A.
ProducerPrice Sample
My goal here was to find producer goods for which one other good is an
important input. Importance is measured by the input's share in the value
24 The path to an answer is strewnwith other difficult questions: Is an input cost
index the appropriate general analogue to the important single input for my sample? Or are there many separate asymmetricresponses to the prices of individual
inputs? For many goods, the single most important input will be labor. Does the
paucity of nominal wage reductions preclude answers to the preceding questions?
etc.
PRICES
RISE
495
of output. For example, cattle are an important input in the production of
beef. For this sample, both the input and output are producer goods whose
prices are measured by a BLS producer price index.
I tried to match inputs and outputs at the least aggregated level permitted
by the data. The PPI is gradually moving toward industry price indexes
based on the standard industrial classification, but the bulk of historical PPI
data are commodity price indexes. I used these commodity price indexes
exclusively. They are classified by commodity codes unique to the PPI. As
with the SIC, PPI commodity codes add digits to reflect more disaggregation. The highest level of aggregation is the two-digit code (e.g., 01-farm
products). The finest level is the eight-digit code (e.g., 01310111-choice
steers) .
Whenever possible, I matched an eight-digit input to an eight-digit output. (For example, I could match the input choice steers to the output
index 02210102-USDA
choice beef carcasses.) However, the available data
frequently dictated the use of six-digit PPIs or, less frequently, four-digit
PPIs (in the preceding example, they would be 022101 -beef and veal and
0221 -meat). Three-digit PPIs were used only to fill gaps in the more disaggregated indexes as described later. Occasionally I combined PPIs to fit the
facts. (For example, urea-formaldehyde plastic resins are made of urea and
formaldehyde. So I combined PPIs for these two components using 1987
cost weights to provide the input index for the resin.)
I began the search for input-output matches with the "use" table of the
1987Input-Output Accounts of the U.S. Economy.This is a matrix based on the
SIC. I used the 1987 version because it provides data for the approximate
midpoint of my sample period. Each column of the matrix gives the share
of one industry's shipments accounted for by purchases from each of the
other (row) industries. These industries are at roughly the three- to fourdigit SIC level, and on average each input-output industry produces around
10 of the eight-digit PPI commodities. From this input-output table, I selected all the entries in which purchases from the input industry accounted
for at least 20 percent of the value of the output industry's shipments (and
both were commodities-producing industries). Thus I eventually matched
choice steers to choice beef because the input-output table first told me
that purchases of "meat animals" account for .763 of the value of the shipments of "meatpacking plants."
The road from the crude matches provided by the input-output table to
the matches eventually used was provided primarily by the Census of Manufactures. This is also based on the SIC, and it provides the source material
for the input-output table. In addition, I used a mapping of PPI commodity
codes into SIC industries provided to me by the BLS.
Table 6A of the census's industry statistics gives disaggregated data on
the value of the products shipped (usually at the seven-digit SIC level) for
each four-digit SIC. Table 7 in the census gives similarly disaggregated expenditures by the industry on materials. The BLS-provided PPI-SIC mapping lists the PPI codes for all commodities produced by each four-digit
SIC industry. This enabled me to match the table 6A output information
and the table 7 input data to specific PPIs. To illustrate with the running
example, the census table 6A for SIC 2011 -meatpacking plants gives the
496
JOURNAL
OF POLITICAL
ECONOMY
value of whole carcass beef (as well as pork, primal cuts, etc.) shipped by
this industry. Table 7Ayields the industry's spending on SIC 013913-cattle
etc.). I then used the SIC-PPI mapping to trans(as well as 013933-hogs,
late the preceding into the specific PPI output-input match: 02210102slaughter steers and
USDA choice beef carcasses/01310111-choice
heifers.
Handling these data requires some irreducible use ofjudgment, common
sense, and general knowledge. For example, the researcher who does not
know that beef is produced from cattle rather than hogs could go astray.
For two groups of products, the matches came from information provided
by industry experts. Roger Fisher, formerly of Amoco Chemicals, provided
information on the technology of producing plastic resins from basic organic chemicals. Donald Barnett, a steel industry consultant; F. M. Scherer,
a renowned industry scholar; andJ. Upton Hudson of U.S. Steel provided
similar information on steel fabrication. The full sample undoubtedly contains some poorly matched inputs and outputs. However, the matching process preceded any data analysis, and I did not discard matches ex post
because of apparently diverse price movements. A detailed listing of the
input-output matches used in the study is available on request.
For each match I sought monthly data from January 1978 through June
1996. The start of the sample was dictated by revision of the BLS's methodology to emphasize transaction prices.25 Some series contain one or more
gaps. I usually filled these gaps from changes in the series at the next level
of aggregation (adjusted for differences in drift).26 Thus, for example, I
beef with data from 022101would try to fill a gap in 02210102-choice
When a four-digit
beef. If that failed, I would use data from 0221-meat.
series was unavailable and the gap was under six months, I filled the gap
with log-linear interpolation. I treated longer gaps as missing values. I
dropped pairs if either series had less than five years of usable data.
I also dropped pairs if there were fewer than 20 input price changes. But
this occurred very rarely. (These cases are in the detailed sample listing
available on request.)
For each input-output pair, I estimated the fraction of 1987 output value
accounted for by the input's cost. The "default" estimate was the one pro-
25See Bureau of Labor Statistics (1988, p. 126). Until 1978, PPI data were often
unrealistically rigid because they reflected "list" prices rather than prices of actual
transactions. A few of the series in my sample still look like list prices-a few changes
the vast majority do not.
punctuating long flat periods-but
26 Specially, let 7h,equal the estimated change in the log of a series with missing
values in month t, t = 1, . . ., T; at, equal the actual change in the log of a related
series without missing values; and M and A equal the levels of the logs of the two
series. My estimate was
7ht
=
at
+
[1
T+ 1
(MT+I - MO) -
(AT+, - AO)].
The last term on the right-hand side is a constant equal to the average monthly
difference in the change of the series I am estimating and the series providing the
estimate.
PRICES
RISE
497
vided by the input-output table. Where possible, I used data in tables 6A
and 7A of the Census of Manufactures to refine this estimate. Thus, instead
of input-output data on meat and meat animals, I used spending on cattle
per dollar of beef shipments for any beef output, spending on hogs per
dollar of pork shipments was used for any pork output, and so forth. For
plastic resins, I estimated the cost shares directly from 1987 prices and the
underlying chemistry provided to me by Roger Fisher. In industries with
substantial interplant transfers (e.g., paper, steel, and nonferrous metals),
the cost share includes these transfers if the census reports their dollar
value.
B.
ConsumerPrice Index Sample
I sought to match a CPI for a specific product with the PPI for the same
product. The main problem here is that the CPI and PPI do not share a
common classification scheme, and there is no off-the-shelf translation
from one classification system to the other. Accordingly, it was necessary to
proceed product by product, with heavy reliance on product descriptions.
While detailed exegesis of the difference between CPI and PPI classifications is unnecessary here, a relevant "bottom-line" difference is that the
CPI product definition is usually broader than the PPI's eight-digit commodity class. Specifically, there are around 450 six-digit PPIs (and more
than this at the eight-digit level) for finished consumer goods. By contrast,
the finest disaggregation of the CPI yields only about 100 consumer goods
indexes. So it is often necessary to use the more aggregated (e.g., fourdigit) PPIs or to combine PPIs to obtain a match to a CPI index.
Another relevant difference is that the CPI classification system is not
conceptually based on specific commodities or items. Instead it grows out
of "expenditure classes" that are used to classify data from the consumer
expenditure survey. The result is a five-digit system with expenditures aggregated into 69 two-digit expenditure classes. Each expenditure class is disaggregated into four-digit "item strata," each of which contains five-digit
bever"entry-level items" (ELI).27 Thus expenditure class 20-alcoholic
ages includes 2005-alcoholic
beverages away from home, which includes
20051 -beer, ale, and other alcoholic malt beverages away from home. Often the item strata and ELIs coincide. So 2001 equals 20011 (beer, ale, and
other alcoholic malt beverages at home). But with some exceptions (called
"special series"), the least aggregated indexes are at the four-digit item
strata level.
As the beer example suggests, even the five-digit ELIs are rather broad
categories. So what happens in effect is that the BLS field force ends up
pricing draft beer in a Philadelphia bar, a bottle of beer in a Chicago restaurant, and so forth to obtain the 20051 component of the 2005 index. Moreover, the identity of the specific products priced at a field office can change
over time. Thus the least aggregated CPI indexes typically include a changing variety of goods within an expenditure class.
27
See Bureau of Labor Statistics (1988) for further details.
498
JOURNAL
OF POLITICAL
ECONOMY
These conceptual differences between the CPIs and PPIs mean that only
approximate matches between them are possible. Moreover, even if we can
match the goods within a CPI/expenditure class index with a PPI, the
weights are not likely to match. In the preceding example, we can match
beverages away from home to the fourexpenditure class 2005-alcoholic
beverages. But this PPI is an aggregate of 10
digit PPI 0261-alcoholic
eight-digit PPIs (bottled beer, sparkling wines, etc.) and three six-digit PPIs
(malt beverages, distilled spirits, and wines). The weights used to aggregate
these detailed PPIs come from the value of factory shipments, not from
consumer expenditures in bars and restaurants.
Finally, there is a potential "slippage" in the match due to transactions
with wholesalers. The PPI measures changes in ex-factory prices, which
are not necessarily the same changes faced by a retailer that buys from a
wholesaler. The same disjunction potentially arises in the producer price
sample too. However, producer-to-producer transactions predominate in
that sample.
These caveats understood, I matched CPIs and PPIs as follows: I began
with the least aggregated CPI index available; usually this was a four-digit
index or one of the special series. Then I sought the closest corresponding
PPI or group of PPIs. This usually meant using some aggregate of eightdigit PPIs. Where feasible, I aggregated eight-digit PPIs (using PPI weights)
instead of using the corresponding six- or four-digit PPI. For example, for
the CPI for living room chairs and tables, I aggregated the eight-digit table
PPI and the eight-digit chair PPI. In this way extraneous items in the sixdigit living room furniture PPI (e.g., desks, cabinets) were excluded. Otherwise, six- or four-digit PPIs were used or, sometimes, combined. I dropped
items from the sample if they required adding PPIs from different two-digit
commodity codes to obtain a match. For example, I dropped the CPI for
sport vehicles and bicycles, which includes outboard motors, boats, and bicycles.28 Here, constructing a corresponding PPI is feasible, but this would
entail too much aggregation for the present purpose. I also dropped CPIs
for which PPI data were too skimpy (because of the coverage of the available
PPIs or gaps in the data).
As with the PPI sample, the construction of the CPI sample preceded
analysis of the data, and a detailed list of the 77 CPI items included and
the corresponding PPIs is available on request.
C. SupermarketPrices
The second-largest supermarket chain in the Chicago area, Dominick's, has
been supplying weekly price and cost data to the University of Chicago's
Graduate School of Business since September 1989. The data are at the
level of the universal product code, that is, a specific brand/package size
such as an 18-ounce box of Kellogg's Corn Flakes. They include all UPCs
in 25 product categories constituting around one-third of storewide sales.
28 The PPIsfor these items can be found under the two-digitcodes for machinery,
transportationequipment, and miscellaneous, respectively.
PRICES
499
RISE
Only packaged goods are included, so there are none of the more volatile
meat, produce, and dairy items.
There are nearly 100 stores in the chain, and every week each store supplies the retail price for each UPC in the sample. Each store is in one of
four pricing zones. These zones are defined by expected long-run pricing
levels based on a combination of costs and competitive constraints. In three
zones (high-, medium-, and low-price zones) the chief competitive constraint is provided by the largest chain in the Chicago area (Jewel). In the
fourth zone, prices are constrained by a discount chain (Cub Foods). Prices
for specific items can also vary among stores within a zone depending on
local competitive circumstances.
In addition to the retail price of the UPC, the database includes the item's
wholesale cost, which is a single chainwide number. This feature makes the
database useful for my purpose. However, the wholesale cost data do not
correspond to those an economist might wish, such as replacement cost or
the last transaction price. Instead, the data pertain to the average acquisition cost (AAC) of the items in inventory.
More precisely, the chain sets retail prices for the next week and also
determines AAC at the end of each week, t, according to
AAC, = inventory bought in t price paid,
+ (inventory, end of t
-
1 - sales,)
AAC,_1.
There are two main sources of discrepancy between replacement cost
and AAC.29The first is the familiar one of sluggish adjustment. A wholesale
price cut today only gradually works itself into AAC as old, higher-priced
inventory is sold off. The second arises from the occasional practice of manufacturers of informing the buyer in advance of an impending temporary
price reduction. This permits the buyer to completely deplete inventory
and then "overstock" at the lower price. In this case, AAC declines precipitously to the lower price and stays there until the large inventory acquired
at that price runs off. Thus the accounting cost shows the low price for
some time after the replacement cost has gone back up.30
For my purpose, the salient problem with using AAC is that it is, in principle, affected by retail price changes. Moreover, the effect is asymmetric.
The endogeneity arises because current retail prices affect sales and thereby
movements in inventory. Thus, for example, a retail price reduction depletes inventory, and this raises the weight on newly acquired inventory in
the calculation of next period's AAC. It can be shown that the combination
of averaging and endogenous weights can induce spurious asymmetries in
the estimated response of retail prices to measured wholesale price.
While caution is therefore warranted in interpreting the results, the measurement problem here needs to be put in perspective. Supermarket inventories turn over more than once per month, and I use data at monthly
29 The following discussion is based on conversationwith MarkMrowiecof Dominick's.
30
Though the path of economic cost here is not so obvious; selling off the large
inventory frees up valuable storage space.
500
JOURNAL
OF POLITICAL
ECONOMY
frequencies. So the typical weight on old prices in my cost estimates is likely
to be small. Moreover, there is evidence that the actual measurement problems are small. They all imply an induced correlation between current retail
price changes and future measured cost changes. (For example, sluggish
adjustment of measured cost means that today's "true cost" will show up
in future measured cost.) Accordingly, in preliminary work I added leads
of wholesale cost changes to regressions including current and lagged cost
changes. The frequency of significant coefficients on the leads was no more
than would be expected by chance.
A change in manufacturer practice led me to terminate the sample period in 1994. Since then, many manufacturers have adopted retrospective
discounts: they announce a discount but deliver it via a rebate based on
how many units of the item were actually sold to consumers in a specified
period. This enables the vendor to limit arbitraging of geographic wholesale price differences. However, the chain's wholesale cost data fail to reflect
these retroactive discounts.
The sample period is generally September 1989 through September
1994. I selected five UPC codes from each product category and collected
their prices at four stores, one from each pricing zone. For each product
category, the stores were selected randomly from within the group of stores
("control stores") not subject to a pricing experiment conducted by the
chain over part of the period. The UPC codes were the five with the largest
sales in the category.
The sample was modified to accommodate the following problems: (1)
For some product categories, it was impossible to identify control stores for
some zones. In these cases, only two or three zones were sampled. (2) There
are gaps in the prices. Their severity varies across categories and, sometimes,
stores. If another store in the same zone sold a UPC at comparable prices,
I used its price, if available, to fill a gap. Otherwise, if the gap was under
two months, I assumed that the retail and wholesale prices remained unchanged. I treated longer gaps as missing values. Most of the longer gaps
occur at the beginning of the sample period because some product categories were added to the database after 1990. (3) Occasionally one of the five
leading items in a category is either introduced or replaced well into the
sample period. I replaced them with the next-best-selling UPC that was sold
continuously in the sample period.
The data are weekly, but I converted them into "monthly" (four-week)
averages. I did this for comparability to the other samples in the study and
to mitigate the aforementioned problems with the use of the AAC. More
important, I used monthly data to reduce the impact of the temporary retail
or wholesale price cuts ("deals") common to the supermarket industry. No
simple time-series model could capture adequately the mix of predictable,
temporary price changes and the more durable, less predictable changes
in the weekly data. The latter are the more interesting for my purpose, and
they predominate in monthly data.
I analyzed two kinds of price series.
1. UPC/store-leveldata.-Here I treat each UPC at each store as a separate
sample. With 25 product categories and up to four stores and five UPCs
PRICES
RISE
501
per category, there are potentially up to 500 such UPC/store samples. I
analyze 357 of them. As mentioned above, I was not always able to obtain
four stores for every product category. In addition, because the goal of the
inquiry is to measure response to price increases and decreases, I had to
drop some UPCs because of an insufficient number of price changes. I
dropped UPCs with fewer than 10 AAC changes in each direction. Because
the data give AAC indirectly (we have retail price and "profit margin" =
[price - AAC]/price), rounding error can induce spurious trivial price
changes. So I adopted a de minimus rule that treats AAC changes of less
than 0.2 percent as no change. Since there are no more than 64 possible
changes per series, my criterion for AAC changes is potentially stringent.
Nevertheless, most UPCs met it. However, the criterion led to elimination
of one entire category (cigarettes), and three (beer, frozen dinners, and
hot cereals) were each reduced to a single UPC by my criterion.
While I sample four stores per UPC, these are not four random draws.
All stores face the same wholesale price, and their retail prices are set by
Dominick's head office. This results in a high correlation between price
changes across pricing zones. Specifically, correlations of price changes for
a UPC between stores in the high, medium, and low zones are on the order
of .8 or .9. This drops to around .6 for pairs including a "Cubfighter" store.
None of the analysis in the paper assumes that the 357 UPC/store samples
are independent of each other.
2. Categoryaverage sample.-For comparability with CPI and PPI data, I
developed a sample in which the product category, rather than the store
or UPC, is the unit of analysis. This was done by averaging across all the
prices within a category as follows: first I converted each price series (Pij,,;
i = UPC, j = store) into an index (Ii j,) with the mean Pij = 100. Then I
averaged the Iij,t across the j stores for each i yielding a UPC index (Ii ).
The category index is then the simple average of the Ii,, across the i UPCs
in the category. This procedure gives each UPC the same weight in the
index.
This sample contains 16 categories as compared to the 24 in the store/
UPC sample. The three categories with a single UPC were dropped. Another five were dropped because the averaging left their AAC index with
too many de minimus price changes.
The categories in the store/UPC sample were as follows (an asterisk denotes a category that does not appear in the category average sample): food
items: ready-to-eat cereals,* hot cereals,* canned soup,* tuna, cheese, cookies, crackers, snacks, "front-end" candy,* frozen dinners,* and frozen entrees; beverage items: beer,* bottled juice, frozen juice, refrigerated juice,
and soft drinks; cleaning/paper: dish detergent, laundry detergent, fabric
softener, paper towels, and toilet tissue; drugs/toiletries: analgesics, toothbrush, and toothpaste.
References
Ball, Laurence, and Mankiw, N. Gregory. "Asymmetric Price Adjustment
and Economic Fluctuations." Econ. J. 104 (March 1994): 247-61.
502
JOURNAL
OF POLITICAL
ECONOMY
Blinder, Alan S. "On Sticky Prices: Academic Theories Meet the Real
World." In MonetaryPolicy, edited by N. Gregory Mankiw. Chicago: Univ.
Chicago Press, 1994.
Blinder, Alan S.; Canetti, Elie R.; Lebow, David E.; and Rudd, Jeremy B.
Asking about Prices: A New Approach to Understanding Price Stickiness. New
York: Sage Found., 1998.
Borenstein, Severin; Cameron, A. Colin; and Gilbert, Richard. "Do Gasoline Prices Respond Asymmetrically to Crude Oil Price Changes?" QJ.E.
112 (February 1997): 305-39.
Buckle, Robert, and Carlson,John. "Inflation and Asymmetric Price Adjustment." Working Paper no. 96-013. West Lafayette, Ind.: Purdue Univ.,
Krannert Grad. School Management, Center Internat. Bus. Education
and Res., 1996.
Bureau of Labor Statistics. Handbook of Methods.Bull. no. 1285. Washington:
Bur. Labor Statis., 1988.
Jackson, William E., III. "Market Structure and the Speed of Price Adjustments: Evidence of Non-monotonicity." Rev. Indus. Organization 12 (February 1997): 37-57.
Karrenbrock,Jeffrey D. "The Behavior of Retail Gasoline Prices: Symmetric
or Not?" Fed. ReserveBank St. Louis Rev. 73 (July/August 1991): 19-29.
Neumark, David, and Sharpe, Steven A. "Market Structure and the Nature
of Price Rigidity: Evidence from the Market for Consumer Deposits."
QJ.E. 107 (May 1992): 657-80.
Sheshinski, Eytan, and Weiss, Yoram, eds. Optimal Pricing, Inflation, and the
Cost of Price Adjustment. Cambridge, Mass.: MIT Press, 1993.