THE MORPHOLOGY OF PRICE DISPERSION

• Date: 01 November 2015
• Author: Greg Kaplan,Guido Menzio
• Published date: 01 November 2015
• DOI: http://doi.org/10.1111/iere.12134

INTERNATIONAL ECONOMIC REVIEW

Vol. 56, No. 4, November 2015

THE MORPHOLOGY OF PRICE DISPERSION∗

BYGREG KAPLAN AND GUIDO MENZIO1

Princeton University and NBER, U.S.A.; University of Pennsylvania and NBER, U.S.A.

This article is a study of the shape and structure of the distribution of prices at which an identical good is sold in a

given market and time period. We find that the typical price distribution is symmetric and leptokurtic, with a standard

deviation between 19% and 36%. Only 10% of the variance of prices is due to variation in the expensiveness of the stores

at which a good is sold, whereas the remaining 90% is due, in approximately equal parts, to differences in the average

price of a good across equally expensive stores and to differences in the price of a good across transactions at the same

store. We show that the distribution of prices that households pay for the same bundle of goods is approximately Normal,

with a standard deviation between 9% and 14%. Half of this dispersion is due to differences in the expensiveness of

the stores where households shop, whereas the other half is mostly due to differences in households’ choices of which

goods to purchase at which stores. We find that households with fewer employed members pay lower prices and do so

by visiting a larger number of stores instead of by shopping more frequently.

mor.phol.o.gy (noun)

The branch of biology that deals with the form and structure of organisms without consideration of

function.

1. INTRODUCTION

This article is a systematic study of the morphology of price dispersion, i.e., the shape and

structure of the distribution of prices at which an identical good is sold in a given geographic

market during a given period of time. Our goal is to provide macroeconomists with a set of

stylized facts that can be used to develop and test theories of price dispersion and to calibrate

models that emphasize deviations from the Law of One Price. We use data from the Kilts-

Nielsen Consumer Panel (KNCP) data set, which contains price and quantity information

for more than 300 million transactions by 50,000 households for over 1.4 million goods in 54

geographic markets during the period 2004–9.

Our approach views both data and theory through the lens of a decomposition of the cross-

sectional variance of prices in a given market and time period. The decomposition expresses

each transaction price as the sum of three sources: one that is specific to the store where the

transaction took place, one that is specific to both the store where the transaction took place

and the particular good that was traded, and one that is specific to the transaction itself. Such a

decomposition can feasibly be implemented in the KNCP, since the data contain information on

transactions for the same good at multiple stores, transactions for multiple goods at individual

stores, and multiple transactions for the same good at individual stores. After investigating the

shape of the typical price distribution, we use the decomposition to study three possible reasons

for price dispersion. Does dispersion arise because some stores are more expensive than others

∗Manuscript received February 2014; revised July 2014.

1This work uses data supplied by the Kilts-Nielsen Data Center at the University of Chicago Booth School of

Business. Information on data access and availability is available at http://research.chicagobooth.edu/nielsen. We are

grateful to the Griswold Center for Economic Policy Studies for providing financial support and to Mark Aguiar

and two anonymous referees for helpful suggestions. Please address correspondence to: Greg Kaplan, Department

of Economics, Fisher Hall, Princeton University Princeton, NJ 08544-1021. Phone: 609-258-4000. Fax: 609-258-6419.

E-mail: gkaplan@princeton.edu.

1165

C

(2015) by the Economics Department of the University of Pennsylvania and the Osaka University Institute of Social

and Economic Research Association

1166 KAPLAN AND MENZIO

on average, because the same good is sold at different prices at equally expensive stores, or

because the same good is sold at the same store at different prices on different days?

We also investigate the shape and structure of dispersion in the average price paid by different

households for the same bundle of goods. We do this by constructing household-specific price

indexes, which is possible because the KNCP data link each transaction to a specific household.

We measure the amount of dispersion in these price indexes and show that price dispersion

does not wash out at the household level. We then use our decomposition to study the sources

of price index dispersion across households, asking if some households pay less than others

because they shop at cheaper stores, because they purchase cheaper goods at the same stores,

or because they purchase the same goods at the same stores on different days.

We start to answer these questions by documenting the extent and shape of dispersion in

the prices for an identical good in a given market and quarter. When we define goods by their

Universal Product Code (UPC, i.e., barcode), the average standard deviation of normalized

prices is 19%. Defining goods more broadly leads to more price dispersion. For our broadest

definition, which aggregates into a single good all products that are identical expect for their

brand and size, the average standard deviation of normalized prices is 36%. We find that price

dispersion is a prevalent phenomenon: While the amount of price dispersion does vary across

goods, the standard deviation of prices is greater than 10% for more than 90% of goods, markets,

and quarters. Moreover, for all definitions of a good, we find that the typical price distribution

is symmetric, unimodal, and leptokurtic, meaning that the distribution has more mass around

the mean and thicker tails than a Normal distribution with the same mean and variance. This

contrasts with the typical features of the wage distribution, which is well known to be skewed

with a long right tail.

To understand the sources of price dispersion, we decompose each transaction price into three

components: (i) a store component, defined as the average price of all goods at the particular

store where the transaction took place; (ii) a store-specific good component, defined as the

average price of the particular good at that store relative to the average price of all goods at

that store; and (iii) a transaction component, defined as the price of the good in that particular

transaction relative to the average price of that good at that store. We find that the store

component contributes approximately 10% to the overall variance of prices, the store-specific

good component contributes between 25% and 45% depending on both how broadly we define

a good and the selection criteria that we apply, and the transaction component accounts for the

remaining variance. These findings are striking since they reveal that price dispersion does not

occur primarily because some stores are cheap and some stores are expensive. Instead, price

dispersion arises because, even among equally expensive stores, the average price of a specific

good varies substantially, and even at a given store, the price of a specific good varies across

transactions.

Our variance decomposition for transaction prices offers new perspectives on the relative

importance of four dominant theories of price dispersion: (i) heterogeneity in the amenities

offered by different stores; (ii) heterogeneity in the marginal costs faced by different stores;

(iii) intertemporal price discrimination; and (iv) search frictions. Deeper understanding about

each theory’s relevance comes from comparing the predictions of each theory with the weight

of each of the three components in our decomposition. Differences in amenities and differences

in marginal costs are both unlikely to be important sources of price dispersion, since these

differences would show up in the store component of prices, which accounts for only 10%

of the overall variance of prices. In contrast, intertemporal price discrimination may be a

quantitatively important source of price dispersion, since the time-variation in prices that stores

use to discriminate among different types of buyers would be reflected in the variance of the

transaction component of prices; this accounts for at least 45% of the overall variance of prices.

Similarly, search frictions may be a quantitatively important source of price dispersion, since

these frictions induce stores to randomize over prices, leading to variation in both the store and

the store-specific good components of prices. Together, these two components account for at

least 35% of the overall variance of prices.

THE MORPHOLOGY OF PRICE DISPERSION 1167

We then ask whether some households are better than others at taking advantage of price

dispersion and, if so, how they achieve this in their shopping behavior. For every household, we

construct a price index by computing the ratio between the household’ s expenditures and the

counterfactual expenditures had they purchased each good at its average price. We find that

the distribution of price indexes in a given market and quarter is approximately Normal with a

standard deviation between 9% and 14%, depending on how a good is defined. Hence, some

of the dispersion in transaction prices disappears when goods are aggregated into bundles, but

a substantial amount of variation remains. Indeed, a household at the 90th percentile of the

price index distribution pays approximately 22% more than a household at the 10th percentile.

Decomposing the variance of household price indexes into a store component, store-specific

good component, and transaction component reveals that the store component accounts for

approximately 50% of the variance of price indexes, whereas the store-specific good component

accounts for 40%, and the transaction component accounts for only 10%.

Three revealing differences about household shopping processes emerge from a comparison

of the variance decomposition for household price indexes with the decomposition for trans-

action prices: (i) there is less dispersion in household price indexes than in transaction prices;

(ii) the store component accounts for a larger fraction of household price index dispersion than

transaction price dispersion; and (iii) the transaction component accounts for a smaller fraction

of household price index dispersion than transaction price dispersion. In Subsection 4.2, we de-

scribe a very simple shopping model that is consistent with these observations. The model relies

on two key assumptions: (i) There is heterogeneity across households in the number of stores

that they regularly visit, and (ii) all households are similar in their abilities to exploit temporary

price reductions by visiting their chosen stores more frequently. We regress household price

indexes on the number of shopping trips taken and the number of stores visited and find that

visiting an additional store has a much larger effect on prices paid than visiting the same store

more frequently, consistent with the simple shopping model that we describe.

We conclude the article by identifying some characteristics of households who pay lower

prices. We find that household price indexes decline monotonically with age: The average price

index for households older than 55 is between 3.5% and 4.5% lower than the average price index

of households younger than 25. We also find that households with more nonemployed members

pay less for the same bundle of goods than households whose members are all employed: The

average price index for nonemployed households is between 1% and 4.5% lower than it is for

employed households. Since older households and nonemployed households are likely to have

a lower opportunity cost of time, these findings support the view that time may be a key input

for the process of finding lower prices.

Our article contributes a systematic analysis of price dispersion to a literature that, until now,

has mostly focused on case studies of particular goods. For example, Sorensen (2000) analyzed

the features of the distribution of prices posted by different pharmacies for several drugs in two

cities in upstate New York. Hong and Shum (2006) documented the distribution of prices posted

by online booksellers for four academic textbooks. Moraga-Gonz´

alez and Wildenbeest (2008)

documented the distribution of prices posted by online sellers for several computer memory

chips. Woodward and Hall (2012) documented price dispersion for mortgage brokerage services.

Galenianos et al. (2012) documented and interpreted price dispersion for several illegal drugs.

Baye et al. (2006) reviewed additional case studies of price dispersion. Although these case

studies are methodologically important, they do not provide a general understanding of the

shape and structure of price dispersion in the retail sector.

A recent and important exception to the lack of large-scale studies of price dispersion is

Eden (2013), who uses transaction data from several supermarket chains in Chicago to relate

the extent of price dispersion to characteristics of goods sold, such as the volatility of demand

for the good, the expensiveness of the good, and the share of revenues associated with sales

of the good. This study is a natural complement to ours, which focuses more on the ‘typical’

morphology of price dispersion.