Change at the Checkout:Tracing the Impact of a Process Innovation

Emek Basker∗

University of Missouri

June 2013

Abstract

Barcode scanners, introduced in the early 1970s, were a foundational processinnovation in the grocery supply chain. By 1984 scanners had been installedin 10% of food stores in the U.S. Difference-in-difference analysis of city-levelprice data shows that scanners reduced prices of groceries by about 1.4% in theirfirst decade. The results are consistent with prior estimates of labor saving byscanners and better information available to stores. Early adopters and adoptersin states that imposed fewer restrictions on complementary process innovationscontributed disproportionately to the price decreases.

JEL Codes: L81, D22, O33

Keywords: Retail, Supermarkets, Prices, Technology, Process Innovation, Bar-code scanners

∗Comments very welcome to emek@missouri.edu. This paper was written while I was a visitor at BostonCollege and at the Stanford Institute for Economic Policy Research; I thank both institutions for theirhospitality. I also thank John Meisel and Sue Wilkinson for help obtaining the Food Marketing Institutereports; C.Y. Choi for help with ACCRA data; Carol Anton, Pat Boyle, Jim Horn, Bill Selmeier, andWalt Waholek for useful industry background; and Ran Abramitzky, David Autor, Jonathan Beck, TimBresnahan, Bengte Evenson, Evangelos Falaris, Cory Koedel, Saul Lach, Ryan Lampe, Mark Lewis, JeffMilyo, Jim Mulligan, Petra Moser, and seminar participants at Arizona, Delaware, Oregon, Stanford, andUSC for helpful comments and conversations. All remaining errors and omissions are my own.

1 Introduction

Process innovations, which allow firms to use the same inputs more productively, are a key

to technological change at both the micro and macro levels. Yet, unlike product innovations,

process innovations are still poorly understood. These innovations are often invisible, as

they may have no observable effects on final products delivered to consumers, but they

represent transformations of firm organization, the production process, and the supply chain.

Differences in process explain why productivity differences persist even within very narrow

industry classifications (Syverson, 2011).

Process differences may be sharpest in the service and retail sectors, where differences

in productivity are particularly striking (Foster, Haltiwanger, and Krizan, 2006). In grocery

retailing, margins are slim yet price dispersion persists. Wal-Mart Supercenters, for example,

charge about 10% less for the same products in the same markets relative to traditional

supermarkets (Basker and Noel, 2009).

This paper provides evidence of the price effect of barcode scanning, a foundational

process innovation in the retail sector. The first scanner came on-line in June 1974 at a Marsh

supermarket in Troy, Ohio. By January 1985, 29% of supermarkets in the U.S. were using the

technology.1 Scanning transformed the grocery sector by allowing supermarkets to encode

prices in machine-readable format and formed the basis for many later improvements in

supply-chain management, from electronic data interchange (EDI) and electronic payments

to inventory management and self-scanning.

Several features of scanning make it a good candidate for studying the impact of process

innovation on prices. First, the Universal Product Code was designed as an industry standard

to preempt individual supermarket chains’ introducing their own proprietary codes. It was

1The diffusion of barcode scanners is covered in a series of papers by Levin, Levin, and Meisel (1985,1987, 1992) and Das, Falaris, and Mulligan (2009). A popular perspective is offered in Brown (1997), and theEuropean background is in Beck, Grajek, and Wey (2011). More generally, the literature on productivity inthe retail sector includes Foster, Haltiwanger, and Krizan (2002), Haskel and Sadun (2009), Doms, Jarmin,and Klimek (2004), Sieling, Friedman, and Dumas (2001), and Lagakos (2013).

hoped that encoding prices in machine-readable form would speed up the checkout process

and reduce its error rate, but it was not designed with consumer prices in mind, and in that

sense represents an exogenous shock to prices. Second, it was a “pure” process innovation,

one that had no impact on the physical product received by consumers. This contrasts with

innovations such as cotton spinning or automobile production lines, which also affected the

final product. Third, the gradual diffusion of scanning across supermarkets provides natural

variation in the proportion of scanning supermarkets across cities and time, which I exploit

to identify the price effect of scanners. Finally, consumer prices of groceries are readily

available and cleanly measured.

I find that grocery prices fell considerably in the first decade of checkout automation. In

a difference-in-difference specification with city-level price data, a scanner penetration rate of

10% — the average across cities at the end of my sample — is associated with grocery-price

decreases of about 1.4%. The largest price effects are for produce and meat, perishable items

over whose prices store managers tend to have the most discretion, and for which scanners

reduced prices by upwards of 2% over this period. Back-of-the-envelope calculations suggest

that the short-run gain in consumer surplus due to scanners is on the order of $7.5 billion

annually in 2012 dollars. I perform several specification tests and “placebo” regressions

to verify that the results are not driven by bias, sample selection, or endogenous scanner

adoption.

One mechanism through which this process innovation affected prices is a reduction in

labor costs. Manual cash registers were simply adding machines; cashiers had to know which

items were taxed and at what rate, and which items were eligible for food stamps. Scanning

was faster than manually entering prices: an early U.S. Department of Agriculture study

predicted that scanners could increase checker speed by as much as 18–19%. In addition,

scanners could be programmed to automatically calculate taxes item by item. These effects

combined to reduce labor costs in scanning stores. In earlier work (Basker, 2012), I found

that scanning stores reduced their wage bills by about 4.5%.

2

Another source of saving is the implementation of complementary processes. Before the

UPC, each can, jar, box, and bag on self-service supermarket shelves had to be individually

tagged with a price sticker. Using the fact that some states required that stores continue

“item pricing” even after installing barcode scanners, while other states allowed stores to

switch to “shelf pricing” — which entailed placing just one sticker on the shelf, rather than

on each item — I find evidence consistent with shelf-scanning providing a significant cost

saving. For items that were not barcoded or tagged with individual price stickers even

before scanning, such as fresh produce, price look-up codes reduced the need for cashiers to

memorize prices.

Additional productivity gains were achieved through changes in worker schedules, in-

creased promotional activity (such as discounting), and the mining of information provided

by scanners. Before scanning, stores did not have product-level data on quantities sold,

and consequently made pricing decisions without accurate estimates of demand functions.

Scanners allowed stores to track sales product-by-product, improve inventory control, and

determine demand as well as obtain better estimates of spoilage and shrinkage rates. This

information allowed stores to adjust prices to demand conditions.

These complementarities can explain why early adopters of scanners generated much

larger price decreases than later adopters, despite an improvement in the technology over

time. Early adopters were likely to be both selected for their technological agility and have

more time over which to make these complementary process adjustments. I estimate that

the marginal effect of an early adopter on average city-level grocery prices is more than twice

the size of the effect of a late adopter of this technology.

The price effect of scanners is larger than the price effects of many other interventions in

the food market. For example, Lach (2007) estimates that a one-percentage-point increase

in the ratio of recent immigrants from the Former Soviet Union to Israel to natives in a

city decreased prices by an average 0.5% due to immigrants’ lower search costs, which drive

down the competitive price level. Basker and Noel (2009) estimate that entry by Wal-Mart

3

Supercenters drives down incumbent supermarkets’ prices in the same city by about 1%.

The rest of the paper is organized as follows. Section 2 provides background on the

technology and the institutional and legal environment. Section 3 describes the data sources

used in this study. Section 4 includes the main results and specification tests. Section 5

evaluates the welfare implications of the finding, and Section 6 concludes.

2 Background

The barcode, or Universal Product Code (UPC), originates with Wallace Flint’s 1932 Har-

vard Master’s thesis. Flint’s idea of encoding price information in punch cards was not com-

mercializable, and went nowhere. In 1949, Norman Joseph Woodland and Bernard Silver

applied for a patent on a similar innovation that was closer to today’s barcode. Implemen-

tation took more than two more decades and an organization, the Ad Hoc Committee on

a Uniform Grocery Product Identification Code, that brought together food manufacturers,

wholesalers, and retailers, with the help of consulting company McKinsey & Co., to agree on

a standard. Food manufacturers in particular were in a hurry to create an industry standard

to preempt individual supermarket chains’ introducing their own proprietary codes, which

would have created a burden on manufacturers.2 By this point, Wallace Flint was a Vice

President of the National Association of Food Chains, and Norman Joseph Woodland was

employed as an engineer at IBM and worked with George Laurer on IBM’s design, which was

eventually selected as the industry standard. The Ad Hoc Committee included representa-

tives of one independent grocer and one cooperative, along with executives of the largest

supermarket chains: Kroger, A&P, and Super Value (Haberman, 2001, p. 145).

The diffusion of scanning took well over a decade, largely because the equipment was

expensive and required a physical reorganization of stores (remodel). Safeway, for example,

2The leader in these early efforts was the supermarket Kroger, working with RCA, which owned theWoodland-Silver patent, on a bull’s eye design.

4

which operated 2,571 stores at the end of 1984, had installed scanners in 925 of its stores by

then, 30% of them in the last year and a similar number in 1983; these scanners were made

by five different manufacturers. Kroger had scanners, made by four different companies, in

approximately half of its 1,318 stores by the end of 1984.3 My conversations with industry

insiders suggest that large chains tended to experiment at first, installing scanners in a few

“representative” stores — for example, stores in different parts of the country or in different

types of neighborhoods (urban, rural, high income, low income, etc.) and added scanners

over time to the rest of the stores. Many chains used multiple suppliers to retain a better

bargaining position. Larger cities with more supermarkets generally saw earlier diffusion, but

holding city size fixed the pattern of adoption across cities does not appear to be systematic.

The early UPC code consisted of 12 digits: a “number system” digit, a 5-digit vendor

identification code, a 5-digit item code, and a check digit. The vendor code was assigned by

a central organization to member companies. Each vendor-member had the ability to assign

up to 100,000 item codes to individual items. As of July 1974, the Uniform Grocery Product

Code Council (UGPCC) had about 1,500 members, most of them food manufacturers; only

members could print the UGPC (later, UPC) on their products, but not every member

printed a barcode on every package, as redesigning package labels and upgrading printers

took time. Absent detailed information on label redesign, it is impossible to know which

packaged products bore barcodes at any given date. For variable-weight products, such as

fresh meat, deli, and some produce, the entire code save the number system and the check

digit was assigned in-store (Selmeier, 2008, pp. 111, 121).

Kroger’s 1975 Annual Report describes the features and benefits of the new technology:

Kroger now has two installations of the new electronic scanner checkout sys-tem [. . . ]. Instead of manually ringing up each price, the checker merely passesproducts bearing an identifying printed bar symbol (Universal Product Code)over a scanner slot in the counter. The scanner sends identification to the com-

3Source: author’s calculations from Food Marketing Institute (various years), Safeway Inc. (1984), TheKroger Co. (1984).

5

puter. . . then the name of the product and the price flash on a screen for theshopper’s view and are printed simultaneously on the receipt tape. Benefits:Faster checkouts, improved accuracy, and a receipt which actually prints out thename of the product as well as the price paid for each item (the tape can beused by the shopper to compare prices from week to week or from store to store).This development represents a major breakthrough in improving checkout ser-vice, and eventually will produce other benefits, such as information recovery formerchandising and for inventory control.

Scanning stores reduced their labor costs by approximately 4.5% (Basker, 2012). There

are two main channels that could have generated this effect: labor saving at the cash register,

due to faster checkout times and/or the ability to use lower-skilled workers (or reduced

training time for new cashiers); and labor saving for price changes on the store floor.

Several features of scanning may make them faster than manual price entry, especially

after a bit of practice. The UPC code is printed in a fixed location on the product label,

and saves looking at the product for the price sticker. Even for items that may not have

been barcoded early on, such as variable-weight products (produce, meat), a fixed code

that could be memorized saved the cashier the need to memorize new prices each time the

price was changed. A store manager remarked to Progressive Grocer magazine in 1976 that

“today the turnover on the front end is so fierce that training the girls [sic] to tell a Rome

from a Delicious apple, or a Texas from a Florida grapefruit, is time consuming. Worse, it’s

pretty ineffective. On the other hand, they quickly memorize most of the codes” (Progressive

Grocer, 1976, pp. 40-42). U.S. Department of Agriculture predictions, reported by Bloom

(1972, pp. 218-220), indicated that if 100% of items were to be barcoded, checker speed

could increase by as much as 18–19%; the productivity gain would have been lower in the

early years, with only partial adoption, but could still have been substantial. Newspaper

reports, as well as informal interviews I conducted with cashiers and store managers about

their personal experiences with scanner adoption, confirm that staffing levels were reduced

in scanning stores (see, e.g., Williams, 1984).

In addition, some scanning stores were able to use the new technology to eliminate the

need to place a price sticker on each and every item in stock. To get a sense of the magnitude

6

of this task, Kroger estimated in 1980 that it was price marking 4.5 million individual items

each year in each of its stores (The Kroger Co., 1980, p. 12). The availability of scannable

UPCs, and of time-invariant codes for variable-weight products, meant that price changes

no longer required workers to remove, and re-apply, price stickers. A second consequence of

this automation was that scanning stores were able to change prices more frequently.4

Scanner-generated data could be used for scheduling, pricing, and inventory manage-

ment. These took time to implement and, with the exception of a small number of highly

savvy retailers, are still work in progress forty years after scanners’ introduction. McK-

insey reported in a 1977 presentation about a Ralph’s grocery store in Los Angeles that

discovered, though the use of scanner data, that its demand curve for generic orange juice

appeared to slope upward for a certain price range (Selmeier, 2008, p. 235). Information on

demand could be used to learn about price elasticities and fine-tune optimal prices (Naka-

mura, 1999). However, many stores and chains did not develop the capabilities to exploit this

information. Inventory control and management is another area in which the potential gains

from additional data were large, yet stock-out rates remain high, and highly variable, across

stores (Matsa, 2011). In many ways the extent to which the technology is being exploited

for maximal efficiency gains still separates the top firms from their competitors.

3 Data

Data on scanner installations come from the Food Marketing Institute (FMI) publication

Scanning Installation Up-Date and provide the month and year of installation by store, from

the first installation of a National Cash Register (coincidentally, the same company that had

produced the first mechanical cash register) scanner at a Marsh supermarket in Troy, Ohio,

4The best estimates of the cost of price changes come from Levy, Bergen, Dutta, and Venable (1997), whofind that price at changes cost (in 1991-1992 dollars) approximately 52 cents per price change in supermarketsnot subject to IPLs, and $1.33 in a supermarket chain subject to such a law. What this implies is that itempricing costs as much as 81 cents per product in 1991-1992 dollars, or $1.33 in 2012 dollars. These highercosts are offset by much lower frequency of price changes, only about 40% of the frequency of non-IPL stores.

7

in 1974 until the end of 1984. The data were compiled by FMI through regular phone calls

to scanner manufacturers.5 These data are described in detail in Basker (2012).6

Ideally, these scanner installations would be matched to high-frequency pre- and post-

installation price data at the store level. Unfortunately, however, store-level price data

spanning early scanner introductions do not exist, at least not on a large scale.7 Instead, I

use city-level average-price data for the period 1972–1984. These come from the American

Chamber of Commerce Research Association, or ACCRA (today, the Council for Community

and Economic Research, or C2ER), which started collecting prices at several hundred U.S.

cities in 1968 for the purpose of creating a “cost-of-living” index. The data are updated

quarterly, in the first week of each quarter (in January, April, July, and October). Several

stores (usually at least five and up to ten) are surveyed in each city for prices of several well-

specified products, such as “one dozen grade A eggs” or “jar of Gerber’s strained vegetables”

(baby food).8 For this analysis, I use ACCRA’s city-level average prices for several products

5I thank Sue Wilkinson from FMI for describing the collection process to me.6As a rule, the sales managers of scanner manufacturers had every incentive to report their installations,

as these figures were viewed by their competitors and publicly established their market presence. NCR’s1980 annual report cites the FMI figures, suggesting NCR believed them to be accurate (NCR, 1980, p. 5).To spot-check the data, I have reviewed the annual reports of several supermarket chains including Kroger,Winn-Dixie, Safeway, and Supermarket General Corporation for statements regarding scanning. While theseannual reports rarely list specific store locations with scanners, they often note the number of stores currentlyoperating scanners. These numbers generally match exactly the FMI figures. For example, Winn-Dixie in its1979 Annual Report says 107 scanners had been installed in its stores by the end of June 1979 (Winn-DixieStores, Inc., 1979, p. 4), perfectly matching the FMI reports. In other cases, the numbers come very close.Safeway reported 40 installations in its 1979 annual report (Safeway Inc., 1979, p. 11); 43 are listed in theFMI database. SGC lists 84 installations that same year (Supermarkets General Corporation, 1982, p. 2),of which 79 are listed in the FMI reports. In a few cases, especially very early on, annual reports list thelocations of scanning stores. Food Fair’s 1977 Annual Report noted one store that had a scanner installedthat year in a Pantry Pride store in Moorestown, NJ, (Food Fair, Inc., 1977, p. 2); this observation is listedin the FMI data. Overall, despite some small discrepancies, I believe the data to be extremely accurate.

7Most micro-level pricing studies today use scanner data, from sources like AC Nielsen’s HomeScandatabase (e.g., Hausman and Leibtag, 2007), Information Resources, Inc. (e.g., Krishnamurthi and Raj,1991), or specific retailers that have shared store-level data, such as Dominick’s Finer Foods (e.g., Chevalier,Kashyap, and Rossi, 2003) or other retailers (e.g., Einav, Leibtag, and Nevo, 2010; Singh, Hansen, and Blat-tberg, 2006). Unfortunately, these types of sources are useless for the current analysis, which requires dataspanning the very introductions of scanners. Store-level prices collected by the Bureau of Labor Statistics(BLS) for the purposes of calculating the Consumer Price Index (CPI) have been used by some researchers,including Cortes (2008) and Matsa (2011), but available records only go back to the mid-1980s.

8The baby-food product definition changed in 1984, switching from Gerber’s to the cheapest available;for this reason I truncate the baby-food price series at 1983q4.

8

that remained consistent throughout the period 1972–1984.9 Stores in the ACCRA sample

are generally chain supermarkets, rather than specialists or limited-assortment grocers. Some

cities in the ACCRA database are actually composites of several nearby cities, as for example

Reno-Sparks, NV or Benton Harbor-St Joseph, MI; but most are stand-alone cities.

The products are listed in Table 1, by general product category. “Non-grocery” items

are items sold either exclusively (as in the case of gasoline) or largely in non-grocery stores

(liquor stores, drug stores, general-merchandise stores, and the like); these non-grocery items

should be impacted minimally, if at all, by the introduction of scanners in grocery stores.

Prices range from an average of 64 cents for a jar of Gerber’s baby food to $19.55 for a

bottle of whiskey, all in constant December 2012 dollars (deflated using the monthly all-items

CPI). Price series (mean and inter-quartile range) for six products with varying degrees of

cross-sectional variance and time-series volatility are shown in Figure 1. Lettuce prices are

particularly volatile due to its short growing season and susceptibility to frost; other prices

are more stable, or trending, over time.

To match the aggregation levels of the scanners and prices, I aggregate the scanner

numbers by city and create a panel of cities with the installed scanner base and prices at the

beginning of each quarter through the end of 1984. I supplement this panel with data from

the Census of Retail Trade (CRT), which provides the number of food-selling establishments

(SIC 54, including supermarkets and smaller grocers, specialists, and other food sellers) by

city for cities with 2,500 or more inhabitants. The CRT is conducted in 1972, 1977, 1982,

and 1987; for intermediate years, I linearly extrapolate the number of stores.

The sample of cities is restricted by two criteria. First, a city has to appear in the

9City-level ACCRA data have been used for a variety of economic studies in the past, including studies ofsupermarket financing (Chevalier, 1995; Chevalier and Scharfstein, 1996); price convergence and deviationsfrom the “law of one price” (Parsley and Wei, 1996; Choi and Wu, 2012); inequality (Frankel and Gould,2001); price pass-through of minimum-wage increases (Aaronson, 2001); and the impact of retailer entryon prices (Basker, 2005; Courtemanche and Carden, 2012). To my knowledge, only one paper has usedstore-level data from ACCRA (Basker and Noel, 2009); the organization has not kept records of store-leveldata from the 1970s and 1980s.

9

ACCRA data set in at least 20 quarters between 1974q1 and 1984q4. Second, the city has to

be listed in the published CRT each year. This leaves 185 cities in 41 states, with the number

of food-selling establishments ranging from 5 (Vermillion, SD, in 1982) to 3,764 (Chicago, in

1972). Sixteen cities appear in the data in each of the 44 quarters from 1974q1 to 1984q4,

and sixteen others appear in the data in just 20 quarters after 1974q1. The average number

of quarters a city appears in the data after 1974q1 is 34.

In addition to the number of SIC 54 establishments, the CRT also provides city-level

counts of the number of SIC 541 establishments and the number of such establishments with

payroll. The first restriction, which limits the count to supermarkets and grocery stores,

eliminates specialty stores such as green grocers and fish mongers but retains convenience

stores and other small general grocers. The second restriction eliminates small family oper-

ations that are staffed by the proprietors and close family members and do not pay wages.

The number of SIC 541 establishments with payroll is only available for cities with 500 or

more business establishments (350 or more in 1987). Of the 185 cities in my sample, 140

have the number of SIC 541 establishments with payroll at some point between 1972 and

1987, and 122 have this information for each Census year. On average, the number of SIC

541 establishments with payroll accounts for about 61% of all SIC 54 establishments in a

city, but this ratio ranges from 25% (for the combined Newark-Elizabeth-Jersey City, NJ

area, in 1982) to 88% (for Augusta, GA, in 1987). This ratio increased over the period of

study, from approximately 55% to approximately 71%, as the number of specialists declined

and larger “combination” meat-and-grocery stores took over the sector. The number of SIC

541 establishments with payroll in the sample ranges from 15 (Fort Collins, CO, in 1972) to

1,362 (Chicago, IL, in 1972).

Of the approximately 150,000 food-selling establishments in the U.S. in 1977, 29,000, or

19%, were in the sample cities. Similarly, of the more than 11,000 scanners installed in U.S.

grocery stores by the end of 1984, 2,332, or about 20%, were in the sample cities. Figure 2

shows the total number of scanners installed in the U.S. over time (dashed line, left axis)

10

and the total number of scanners installed in the sample cities (solid line, right axis). The

earliest scanner installations in this sample of 185 cities were in Houston and Indianapolis,

both during the second quarter of 1975. Fifteen cities in the sample still did not have a

single scanning store by the end of 1984. By the end of 1984, the largest number of scanners

(225) was installed in Houston, and the largest number of scanner as a fraction of either the

number of food-selling establishments or the number of grocery stores with payroll was in

Eau Claire, WI (52% or 85%, respectively). City-level summary statistics are in Table 2.

I supplement the city-level analysis with MSA-level analysis using Consumer Price Index

(CPI) data. Until 1997, the BLS published monthly MSA-level price indices for food at home

as well as five sub-categories of food at home: fruits and vegetables (produce); dairy products;

meats, poultry, fish, and eggs (meat); cereals and bakery products (bakery); and other food

at home. Unlike the ACCRA data, the indices do not capture level differences in the prices

at different MSAs, and all are normalized to 100 in 1982–1984. These data are available

for 27 metropolitan areas.10 Tables 3 and 4 list, respectively, the metropolitan areas and

the products for which CPI data are available. Aggregating the scanner installations to the

MSA level, the 27 MSAs in the data account for nearly 44% of scanner installations.

4 Scanners and Prices

4.1 Difference-in-Difference Estimates

To determine the average effect of scanners on city-level prices, I estimate the following

difference-in-difference equation on the set of 15 grocery products:

ln(price)ijt = αij + δjt + θScanFracit + εijt (1)

10The “meat” and “other” CPIs start in 1976; the rest are available for the entire time period of thisstudy, with the exception of all series in the Miami-Ft. Lauderdale, FL metro area, which start in 1977.Most CPIs are monthly, but the San Francisco CPI skipped December 1973 and the Denver-Boulder CPIwas quarterly from April 1972 to January 1977, after which it became monthly.

11

where priceijt is the average price of product j in city i in the first week of quarter t,

αij is a city-product fixed effect, δjt is a product-time fixed effect, and ScanFracit ≡(Scannersit/SIC54it

)is the number of scanners installed in city i at the beginning of quarter

t, scaled by the number of food-selling establishments in that city in the that year. Standard

errors εijt are clustered at the city level to allow for arbitrary autocorrelation in the error

term as well as arbitrary correlations in the error term across products within a city, both

contemporaneously and across time periods.

The key identification assumption in Equation (1) is that prices in cities with more/earlier

scanner installations and cities with fewer/later installations would have received i.i.d. shocks

but for the scanners; put differently, if the Ad-Hoc Committee had not agreed on a symbol

in 1973 and scanning had not been possible, no systematic differences would have been seen

between the price changes in early-adopting cities and those in late-adopting cities.

There are some data-quality concerns for the price data, both because of aggregation

issues and because the data are collected by employees and volunteers who do not have special

training in survey methodology. Although I do not have information about the instructions

given to price collectors during the period of study, in recent years ACCRA has supplied

a detailed instructions manual to price collectors, including the types of stores they should

survey (stores where professionals shop), the number of stores (at least five), what to do in

the event the product is out of stock, etc. Perhaps more importantly, because prices are

a left-hand side variable in this analysis, any i.i.d. errors should be incorporated into εijt,

and any systematic bias at the product, city, or product×city level will be captured by the

fixed effects αij. The only remaining concern, then, is that systematic errors or biases in

price collection were correlated with the introduction or diffusion of scanners; this concern

is addressed in several specification tests in Section 4.3.

Pooling all observations into a single regression, the point estimate for θ in the pooled

regression is approximately −0.14, statistically significant at conventional levels. This figure

is shown in the first column of Table 5. To interpret this coefficient, note that the average

12

value of ScanFrac in the data, in the last quarter in which cities show up in my dataset,

is 10%. Extrapolating the effect of scanners to the case where ScanFrac approaches one is

far out of sample. Many of the establishments included in SIC 54 are not the sort of large

grocery stores and supermarkets that installed scanners in the early years; in the 1977 CRT,

for example, 30% of SIC 54 stores were specialists including meat and fish markets and

bakeries, and another unknown fraction were convenience stores and other small grocers.

Instead, I evaluate the estimates at ScanFrac = 10%. The product 0.10 · θ provides an

estimate of the short-run impact of the fraction of scanning grocery stores in the city within

the range for which data are available: an average price decrease of 1.4%.

To get a sense of the magnitude of these figures, if demand at an individual supermarket

has a constant elasticity ε, then price is a constant markup over marginal cost(p = ε

ε−1c),

and the marginal-cost elasticity of price(

dpdc

cp

)is 1, so a 1.4% decrease in price implies

a 1.4% decrease in marginal cost for the average store. For stores that installed scanners,

Basker (2012) finds that scanners reduced store-level labor costs by 4.5% on average. Official

industry-level estimates of the fraction of costs accounted for by labor are not available, but

conversations with industry insiders and anecdotal observations from supermarket chains’

annual reports suggest that labor costs were generally below 15% of costs during this time

period. If this is accurate, and if the 4.5% decrease in total labor costs found in Basker

(2012) also translates into a 4.5% decrease in marginal labor cost, then marginal cost fell by

no more than 0.045 · 0.15 ≈ 0.67%. Since the price decrease I estimate is twice that size,

this calculation suggests that either marginal cost fell by more than average cost, or stores

were able to use the information generated by scanners to refine their pricing, bringing prices

closer to the optimal markup over marginal cost.

To better understand the mechanism for this effect, I next test for heterogeneous price

effects over three margins: products, time, and regulatory environments.

13

4.2 Heterogeneity

4.2.1 Product-Level Heterogeneity

Supermarkets had varying degrees of control over pricing of different products. For many

packaged goods such as canned foods, prices were negotiated with suppliers well in advance

and covered wide geographic areas; negotiations included the timing, frequency, and level of

any temporary discounts (Marr, 1996). Other products, such as milk, were subject to price

controls in many areas. Conversations with industry insiders suggest that local supermarkets’

pricing power was greatest for locally sourced foods, including produce, meat, and bread.

In contrast, the biggest scanning-induced savings were likely to have accrued to the

nationally sourced packaged goods that were early adopters of barcodes. The exact dates

at which individual products started featuring barcodes are not available, but we do know

that several prominent canned-goods producers were early members of the UGPCC. Hunt-

Wesson, manufacturer of Hunt’s canned tomatoes, had three separate UGPCC registrations

as of July 1974, as did Del Monte, another large manufacturer of canned goods; Nestle (which

had acquired Libby’s, a canned-food manufacturer, in 1971) had four registrations (Distribu-

tion Codes, Inc., 1974). Each registration allowed a company to print up to 100,000 separate

UPC symbols, so three registrations would have covered as many as 300,000 products. It

seems likely that for most, if not all, the period of this study, supermarket scanners would

have been able to scan canned tomatoes and similar products.

Scanners also generated indirect cost saving for the produce and meat departments. In

the case of produce — bananas, heads of lettuce, bags of potatoes — the products generally

did not have individual price tags prior to scanning and did not get barcoded when stores

introduced scanners. In fact, produce displays changed barely if at all with the introduction

of the new technology. Instead, scanning stores introduced price look-up codes (PLUs) for

most fresh products. These codes saved cashiers the need to memorize an ever-changing list

14

of prices or waste valuable time looking up these prices.11

For meat, scanning stores had the option of installing new scales to complement the scan-

ning technology. Bill Selmeier, an IBM engineer involved in early scanner installations and

maintenance, recalls early meetings with Toledo Scale and Hobart Scale, two manufacturers

of meat scales, to ensure that in-store meat scales would be able to print meat packages with

UPC symbols when scanning systems were installed in stores (Selmeier, 2008, pp. 131-132).

Stores that had these scales were able to barcode variable-weight meat products such as

whole chickens and ground beef.

To get a sense of the degree to which price reductions varied across products, I re-

estimate Equation (1) separately by product category. These results are shown in columns

2–6 of Table 5. Each pooled regression includes three products, by category — produce,

dairy, meat, canned, and miscellaneous grocery items. As above, I allow both the time

fixed effects and the city fixed effects to vary by product, and cluster the standard errors

at the city level to allow for arbitrary correlations in the error term both within and across

products. The coefficients are dramatically different across product category, with the largest

price decreases — corresponding to approximately 2.8% and 2.1% at 10% adoption — for

produce and meat products, followed by miscellaneous products (1.2%) and dairy (0.7%,

not statistically significant at conventional levels), and essentially zero for canned goods.

Repeating this analysis product-by-product, I find that most of the variation in θ occurs

across, rather than within, product categories. To conserve space, the predicted price effects

for each of the 15 products are reported in Figure 3. The length of the bars shows the size

of the short-run effect on prices, and the significance level is denoted by the shade of the bar

(white bars are statistically insignificant at conventional levels). Fourteen of the 15 effects

11PLUs were also available in some non-scanning stores that installed electronic cash registers. The1976 Annual Report of the Great Atlantic and Pacific Tea Company (A&P), for example, reported that“[e]lectronic cash registers capable of being tied into electronic scanners were installed in 81 existing storesand all new, enlarged and remodelled [sic] stores” (The Great Atlantic and Pacific Tea Company, 1976, p. 3).Thus, all scanning stores and a subset of non-scanning stores experienced this PLU effect.

15

are negative; of these, three (bananas, lettuce, and chicken) are significant at the 1% level,

three more (potatoes, bacon, and shortening) are significant at the 5% level, and one (beef)

is significant at the 10% level. The single positive coefficient, not statistically significant, is

for the price of canned tomatoes.

These results suggest that supermarkets engaged in some degree of cross-subsidization,

reducing prices of the products over which they had more control — most immediately, meat

and produce — rather than those that were priced at the regional level in negotiations with

large national manufacturers. To the extent that the prices of canned goods fell due to

scanning, that effect is likely to have been considerably more geographically diffused: when

most or all of a given chain’s stores in a given region or division were scanning, it could

reduce prices of these goods region-wide. These effects may have been large but they cannot

be observed in a city-level difference-in-difference specification.12

4.2.2 Regulatory Heterogeneity

When scanning was still in its infancy, several states — under pressure from unions and

consumer groups — passed so-called “item-pricing” laws (IPLs), requiring stores to retain

item pricing even in the event that technological advancement made this practice obsolete.13

The Massachusetts Attorney General passed an item-pricing regulation as early as 1971, with

the goal “to ensure that check-out clerks were not attempting to charge prices by memory,

and to prevent pricing discrimination between customers” (Hurst, 2007). As of February

1976, three states – California, Connecticut, and Rhode Island — had passed such laws, and

eight more were considering such action (Wilson, 1976); by the end of 1976, Michigan and

New York had also passed IPLs (Das, Falaris, and Mulligan, 2009). Michigan’s law became

effective January 1, 1978 (Anderson Economic Group, 2010).

12A similar result on the higher sensitivity of perishables’ prices to competitive conditions is reported inBinkley and Connor (1998).

13Levy, Bergen, Dutta, and Venable (1997), Bergen, Levy, Ray, Rubin, and Zeliger (2008), and Levy(2008) make a compelling case that IPLs are costly to stores and, indirectly, to consumers.

16

This regulatory heterogeneity affected the degree to which stores were able to comple-

ment the new technology, scanning, with an additional process change, shelf pricing. More

flexibility in the adjustment process is likely to have increased a store’s surplus from scan-

ning, an effect that should be reflected in prices. But the regulations had loopholes, and

even non-regulated stores were constrained in many ways.

On the one hand, some stores and products in IPL states were exempt from item-pricing

regulation. In New York, for example, stores generating less than $3 million annually were

exempt from the item-pricing law; milk, eggs, and unpackaged fresh produce were exempt

in all stores, as were products priced at less than 50 cents; and stores could also choose to

forgo item pricing on up to 3% of their products (Barron, 1984). In Michigan, items sold by

weight or volume were exempt from the law, as were very small items and those priced at

less than 30 cents (McGrayne, 1986). California law exempted items priced at less than 40

cents.

At the same time, although most stores in non-IPL states faced no legal or regulatory

restrictions on item-pricing removal (there were exceptions due to municipal regulation in

some areas), consumers were generally opposed to such changes. As a result, even stores in

non-IPL states often retained item pricing for a long period after scanner adoption. It was

not until 1981, for example, that Giant Foods, Inc. removed item prices in its Washington,

DC area stores; Safeway followed soon after (Richburg, 1981).

To test whether item-pricing laws limited the cost-saving (and price pass-through) of

scanners, I estimate an augmented version of Equation (1) in which I interact the scanner

variable with an indicator for an item-pricing law:

ln(price)ijt = αij + δjt + θ0IIPL=0ScanFracit + θ1IIPL=1ScanFracit + εijt (2)

I let the IPL indicator equal 1 for all products in an IPL state, whether or not those products

were exempt from the law. The rationale for this definition is that, particularly given the

17

evidence in the previous subsection about potential cross-subsidization, a state with item-

pricing laws is one with a different regulatory environment, and this legal environment may

affect the prices of exempt as well as non-exempt products. In addition, exemptions that

are based on prices are clearly problematic due to the endogeneity of the regulation to the

outcome variable of interest.

The results are given in Table 6. As in Table 5, the first column shows the coefficient

estimates pooling all products in one regression, and the next five columns show results for

sets of three products by product category. Overall, scanners lowered prices in non-IPL cities

by an average of 1.6% over this time period (coefficient estimated at ScanFrac = 10%), but

had no effect on prices in IPL cities. A two-sided test for equality of the coefficients rejects

the null with p-value 0.0159.

There is considerable heterogeneity by product in the differential effect of scanners on

stores constrained by IPLs and stores left unconstrained. In particular, while produce prices

in both regulatory regimes fell with the introduction of scanners, prices of dairy and canned

goods (and, to a lesser extent, prices of miscellaneous grocery products) in IPL cities are

estimated to have increased, rather than decreased, with scanning.

Because even stores in regulatory regimes that allowed item-pricing removal did not

generally remove item pricing for some time after installing scanners, these estimates are

a lower bound on the actual effect of item pricing. There is no way to know when stores

in the ACCRA sample stopped the practice of item pricing. One crude way of capturing

this is to interact the scanner variable with an indicator for observations during and after

1981, because item pricing was generally phased out later in the sample. These estimates

(not shown) suggest that the effect of scanners was very similar in IPL and non-IPL cities

before 1981 (point estimates are approximately −0.1 for both regimes, although neither is

statistically different from zero), but a divergence in the effect of scanners in the later period.

18

4.2.3 Selection and Technological Heterogeneity

Although basic scanning technology did not change over the period of study, scanners did

improve over time. Early scanners provided only faster processing of customer purchases and

an itemized receipt. From May 1979 to November 1980, a second vintage diffused; models

installed during this period provided faster service by allowing the scanner to read much

smaller bar codes. A bigger change was the introduction, in November 1980, of a holographic

supermarket terminal, which increased the distance at which scanners could read barcodes

(International Business Machines Corporation, 1980, p. 4). This improvement, which Das,

Falaris, and Mulligan (2009) call “vintage 3,” also read damaged and crinkled barcodes (Das,

Falaris, and Mulligan, 2009), and had some inventory-management capabilities. Stores that

had already installed vintage 1 or vintage 2 scanners, however, did not upgrade to the

improved scanners for several (sometimes many) years, given the expense involved.

As scanner quality increased, prices fell (Zimmerman, 1999), and with them the size of

the stores installing scanners (Basker, 2012). This selection effect implies that the stores

with earlier installations may have had more to gain from scanning, or may have been more

technologically sophisticated, and may consequently have been able to make larger price

reductions due to scanning. In addition, the larger early adopters may have, independently

of the date of adoption, larger catchment areas and therefore a bigger effect on competitors’

prices. Finally, later adopters would have had less time before the end of the price panel to

fully adjust to the technology.

If the improvement in technology dominates the price effect, cities with more late-vintage

adopters should have experienced larger price declines; if the selection effect dominates, cities

with more late-vintage adopters should have experienced smaller price declines. To empiri-

cally assess which effect dominates, I estimate a variant of Equation (1) in which I separately

count scanner installations by vintage. Following Das, Falaris, and Mulligan (2009), I col-

lapse vintages 1 and 2 into a single category due to the small number of installations in

the very early period. Installations from 1974 to April 1979 were “vintage 1” installations;

19

installations between May 1979 and October 1980 were “vintage 2,” and installations from

November 1980 to the end of the sample period are “vintage 3.” I estimate

ln(price)ijt = αij + δjt + θ12ScanFracV.1&2it + θ3ScanFracV.3it + εijt (3)

where ScanFracV.1&2 is the number of scanners of vintages 1 and 2 installed in city i as

of the first week of quarter t, normalized by the number of SIC 54 establishments in that

city that year, and ScanFracV.3 is the number of vintage-3 scanners normalized in the

same way. ScanFracV.1&2, therefore, increases between 1974 and 1980 and remains fixed

thereafter, whereas ScanFracV.3 is zero through 1980 and starts increasing in the first

quarter of 1981. As of the end of the sample period, a third of scanner installations in the

sample cities were of vintages 1 and 2 (the vast majority of these vintage 2), and two thirds

were vintage 3.

The results are shown in Table 7. For the full set of products as well as for every subset

of products, the coefficient θ12 is estimated to be larger in absolute terms (more negative)

than θ3. The differences are significant at the 5% level in the case of canned goods and at

the 10% level for dairy and the full set of products.

These result indicate selection played an important role in the timing of adoption of the

technology, overwhelming the effects of better technology in later years.14

In terms of the overall effect on prices, these results suggest a roughly equal share of the

overall price decline due to scanning is attributed to the second and third vintages. These

results also suggest that scanners’ full effect on prices may be very close to the short-run

effects, if later adopters had a sharply decreasing impact on prices.

These results highlight both the importance of timing and the limitations of the scanning

technology. Timing matters because the effects of the technology are cumulative, due to

14I rule out a simple nonlinear effect by estimating Equation (1) on the set of 57 cities in which there wereno vintage 1 or 2 installations, and still find the smaller point estimates associated with vintage 3 scanners.

20

the adoption of complementary processes such as inventory management practices, worker

training and scheduling, and item-price removal. The limitations of the technology matters

because scanning is only as effective as the stores that adopt it. Scanning did not merely

speed up the checkout process; its potential was far greater. But if supermarkets were slow

to make the complementary process innovations to take full advantage of scanning, actual

gains fell far short of this potential.

4.3 Other Explanations

The identification of the effects of scanners on prices above depends on the standard difference-

in-difference assumption that the timing and extent of treatment (in this case, scanner adop-

tion) is independent of the error term. While I cannot test this assumption directly, I test

for a couple of specific concerns: systematic bias in the constructed variable ScanFrac and

changes in product definitions in the ACCRA data. I then perform two “placebo tests” for

other endogeneity/omitted-variable concerns.

4.3.1 Bias

The variable ScanFrac, as constructed here, is both mismeasured and biased. This is be-

cause the denominator used is the number of SIC 54 establishments in the city as measured

by the Census of Retail Trade. SIC 54 is a very inclusive measure of establishments, in-

cluding both supermarkets and large grocery stores and very small stores, many of them

specialty stores — bakeries, butchers, candy stores, green grocers, and the like. These small

specialty stores are not legitimately “at risk” of scanner installations during the period of

study, nor are they likely to have been included in the ACCRA surveys, except possibly in

very small markets where there were very few, if any, large grocery stores. The over-count

of the denominator shrinks ScanFrac, with several effects. First, downward bias in the

variable causes upward bias in the coefficient estimate θ. Second, the bias is accompanied

by measurement error, which may attenuate the coefficient. Finally, if scanner adoption

21

is correlated with the measurement error in ScanFrac — for example, if small cities have

disproportionately few supermarkets and also tend to have later scanner adoption — this

could further bias θ.

For the smaller sample of (larger) cities for which the number of SIC 541 establishments

with payroll is available, I can estimate a regression with a more precise measure of the

fraction of scanning grocery stores:

ln(price)ijt = αij + δjt + θScanFrac∗it + εijt (4)

where ScanFrac∗it ≡

(Scannersit/SIC541it

)is the number of scanners normalized by the

number of grocery establishments (SIC 541) with positive payroll.

The results are given in Table 8. Consistent with the concerns about bias inflating the

previous coefficient estimates, point estimates of θ are smaller across specifications. However,

because the average end-of-sample value of ScanFrac∗it for cities with at least one scanning

store is about 16%, I now evaluate the short-run effect by multiplying the coefficient estimate

for θ by 16% rather than 10%. The resulting short-run estimates are larger in absolute value

than the prior ones, a symptom of measurement error in the original variable. The overall

pattern of coefficients is the same as in Table 5, but now the estimate for the average effect

of scanning is a third larger: a price decrease of about 2% overall, and as high as 4% for

produce. In addition, the coefficient estimate in the dairy-products regression is significant

at the 10% level (it was insignificant previously), and the coefficient in the miscellaneous-

products regression is significant at the 5% level (tighter than the 10% level previously).

4.3.2 Product and Store Selection

One concern about the ACCRA data is that the products chosen by ACCRA to be easily

comparable across stores and cities may not be representative of the full shopping basket.

Another concern is that most of the product descriptions in Table 1 do not specify a brand.

22

In these cases, the brands may not be the same across stores and cities or within a city over

time. This could cause bias in the above estimates of θ if product substitutions coincided

with, perhaps were even caused by, scanner adoption. (In the latter case, θ becomes an

estimate of product innovation rather than process innovation.) One way this effect could

manifest is through the introduction of store brands (generics). McEnally and Hawes (1984)

and Ward, Shimshack, Perloff, and Harris (2002) point to the late 1970s as the watershed

period for the introduction of store-brand groceries. Kroger reported in its 1975 Annual

Report that 25% of store revenues came from its own store brands (The Kroger Co., 1975).

Later scanner-generated data shows that private labels’ market increased across product

categories in the late 1980s and early 1990s (Hoch, Montgomery, and Park, 2000), but

comparable data spanning the introduction of scanners are lacking.

A related concern is that the list of stores surveyed by ACCRA in any given city may

have changed over time, both due to entry and exit of stores and due to changes in each

surveyor’s assessment of which stores provide representative prices for the local area. This is

because the ACCRA survey is intended to maximize the cross-sectional accuracy, but makes

no claims about time-series consistency. This, too, may bias the estimates of the effect of

scanning if stores that start scanning become more (or less) likely to be included in the

survey.

To test whether store or product substitution and/or selection are driving my results,

I estimate a variant of Equation (1) using metro-area CPI data from the Bureau of Labor

Statistics. Historically, the CPI “linked out” product and outlet substitutions, a practice

that is sometimes controversial because it overstates inflation, but is useful in this context.

The CPI is normalized to 100 for each city, and each product category or sub-category, in

1982–1984, so it does not allow cross-sectional level comparisons, but it still allows difference-

in-difference analysis. For each of the six food-at-home and component price indices, I

estimate

ln(CPI)it = αi + δt + θScanFrac∗it + εit (5)

23

where CPIit is the Consumer Price Index, in levels, for metro area i at month t. Similar

to the city-level regressions, αi is a metro-area fixed effect, δt is a time fixed effect, and

ScanFrac∗it is the number of scanners installed in metro area i at the beginning of month t,

scaled by the number of SIC 541 establishments with payroll recorded for that metro area.

Standard errors εit are clustered at the metro area level.

Results, shown in Table 9, are necessarily imprecise; only 27 metro areas are used to

identify the coefficients. Nevertheless, the results are consistent with the ACCRA regression

results. The first column of Table 9 reports the results for the CPI for food at home; each

of the subsequent columns reports results for a different component CPI. The coefficient on

ScanFrac∗ is systematically negative. As with the city-level price regressions, the coeffi-

cient is largest for meat and produce, is statistically different from zero at the 5% level in

the meat-products regression, and at the 10% level for both the overall food-at-home and

the miscellaneous-products regressions.15 This result adds a measure of confidence in the

ACCRA data’s suitability for panel analysis.

4.3.3 Placebo Tests

In addition to prices of grocery products, ACCRA collects and reports prices of several non-

grocery items. These include items that are rarely sold in grocery stores, like gasoline, and

items sold by both grocery stores and many non-grocery (and hence, in this time period,

non-scanning) stores including general merchandisers and drug stores, such as cigarettes.

I re-estimate Equation (1) using three of these products: cigarettes, gasoline, and alcohol.

According to the 1977 CRT, grocery stores accounted for 42% of all alcohol sales by value,

23% of all tobacco sales, and less than 2% of all automotive fuel and lubricant sales. (By

contrast, grocery stores accounted for 89–93% of sales of dairy products, fresh produce, and

15The Census definitions of many metro areas changed over time. While I have made efforts to adjust thedenominator to create a set of time-invariant MSA definitions, the results do not appear to be sensitive tothese adjustments. Using the number of SIC 541 establishments with payroll listed in the 1977 CRT as thedenominator throughout the sample period yields similar coefficient estimates.

24

meat.) Because early scanners were installed almost exclusively in grocery stores, these

prices may be minimally affected by scanning, but the effects should be much smaller than

that for grocery products.

Estimation results are shown in Table 10. For the three non-grocery products combined,

the coefficient on the scanner variable is very small and tightly estimated. Of the three

individual products, two have negative coefficients and one has a positive coefficient; none

are statistically significant. The largest negative effect is on the price of whiskey. This is the

product whose share of sales in grocery stores is the largest of the three, so we may expect

some causal effect of scanning on this price.

Overall, the results suggest that there are no overall declining price trends that can

provide an alternative explanation for the main results in the paper.

The test also cannot pick up on supermarket-specific endogeneity concerns. Specifically,

if price decreases declined because scanning supermarkets, and/or their competitors, imple-

mented other changes coincident with scanning, estimates of θ could reflect this omitted

variable, such as ECR installations. As long as the two innovations, scanning and the unob-

served variable, do not always occur exactly together, but the omitted innovation sometimes

precedes and sometimes follows (or coincides with) scanning, they can be teased apart in

principle. Table 11 shows the results of an augmented regression in which next quarter’s

scanner adoption rate, ScanFraci,t+1, is included along with the current quarter’s:

ln(price)ijt = αij + δjt + βScanFraci,t+1 + θScanFracit + εijt (6)

The coefficient on the lead of the scanner adoption rate, β, tells us whether prices started

changing immediately prior to an increase in the scanner penetration rate in a city. A signif-

icant coefficient on the lead scanner variable could indicate reverse causality of the estimates

in Table 5. A zero coefficient, in contrast, lends credence to the exogeneity assumption.

Although the standard errors are large, there is no evidence that price decreases preceded

25

or anticipated scanner diffusion; if anything, the opposite is true. The effect on the lead coef-

ficient is positive in the regression with all products and in four of the five product-category

regressions; it is negative only for one category (produce). In contrast, the estimated price

effect of current-period scanning increases (in absolute value) for most product categories.

5 Welfare Implications

The demand for food is quite inelastic, especially outside of the very poorest populations,

so a 1.4% decrease in the price of food is associated with an increase in consumer surplus

amounting to approximately 1.4% of total food expenditure. The 1977 Census of Retail

Trade calculates consumers’ total expenditure on groceries and other food (merchandise line

100) at $137.5 billion in 1977 dollars, or about $510 billion in 2012 dollars. The 1982 Census

of Retail Trade calculates retail expenditures on merchandise line 100 at nearly $200 billion

in current dollars, or $470 billion in 2012 dollars. An increase in consumer surplus equivalent

to 1.4% of food expenditure is therefore on the order of $7.5 billion in 2012 dollars.

This calculation is a lower bound in several ways. First, the ACCRA prices used here do

not account for substitution across stores or products as a result of changes in relative prices.

Stores in the ACCRA sample are weighted equally, so the only substitution these prices allow

are due to store entry/exit from the sample. This, in turn, can occur non-randomly if stores

open or shut down or if stores become less attractive and the surveyor determines that

professionals are no longer likely to shop there. While these effects do introduce some degree

of substitution into the data, the average quantity-weighted price is likely to be much lower

than the average price reported by ACCRA. Whether the quantity-weighted price declines

more or less sharply depends on which stores lower their prices and on the elasticity of

substitution in the demand function.

Second, barcodes were just the first domino in a long chain of innovations that have

transformed the food supply chain, and over time other sectors as well. The current estimates

26

pertain only to the very earliest savings, due to faster checkout times, lower costs of restocking

and price changes, and early inventory management, and do not incorporate improvements

the occurred after 1984. To the extent that barcodes really did lead to frequent deliveries and

superstores (as hypothesized by Holmes, 2001), they have helped chains generate economies

of scale and scope and have had an even larger impact on prices by creating opportunities

for the reorganization of the supply chain. Quantifying these long-run effects is considerably

more challenging due to the intervention of confounding factors, but the effects estimated

in this paper suggest that barcodes induced large effects even in the short run, before these

add-on complementary technologies were introduced.

In addition, the estimate assumes nonlinear price effects so that the full effect is realized

during the sample period (by the mid-1980s), with an adoption rate of 10% on average, and

does not account for measurement error, which would increase the estimate by about one

third.

6 Concluding Remarks

The underlying causes of productivity differentials across firms and establishments have too

often been a black box, assumed to be some combination of physical capital, information

technology, and labor practices. This paper takes a first step towards measuring a process

innovation at the store level and tracing its impact in order to learn about the mechanism

through which it affects productivity and the magnitude of its effect on pricing. If scanners

alone, in the short run, can produce a noticeable price effect in a sector with razor-thin

margins, then technological differences, narrowly defined, may well explain price differences

among retailers.

The analysis in this paper implies that a conservative estimate of the full impact of scan-

ners — before later add-on innovations such as electronic payments, but in conjunction with

27

complementary technologies of the day, specifically price look-up codes and early inventory

management — is a decrease in U.S. grocery prices of at least 1.4% in the short run.

The estimated price effect is large, but consistent with other estimates of the effect of

technological process innovations. Brynjolfsson and Hitt (2003), for example, find supra-

normal returns on other IT investments, up to five times greater than normal returns on

investment. Using data from supermarkets in the New York-New Jersey-Connecticut tri-

state area, where some supermarkets were subject to item-pricing laws and others were not,

Bergen, Levy, Ray, Rubin, and Zeliger (2008) find that prices in supermarkets subject to

IPLs are 8–9.6% higher than prices in non-IPL supermarkets. This last finding may be

partly due to the fact that supermarkets in non-IPL states have made more investments

in complementary cost-lowering technologies and organizational changes, such as expanding

product selection. Selection due to entry and exit may also play a role.

A limitation of the price data used here is that we cannot observe the price distribution

across stores within a city, and therefore cannot determine how disproportionate the price

effect was in scanning stores relative to their competitors. Addressing this question would

require matching store-level price data to the FMI dataset. While this cannot be done on a

large scale, it may be possible to do for one city or a small set of cities using prices reported

in local newspapers or advertised in newspaper circulars.

Estimating heterogeneous effects of scanners on prices by product, regulatory regime,

and time of adoption exposes both the strengths and the limitations of the technology.

First, although supermarkets’ cost savings were spread across a large set of products, price

reductions were concentrated in a relatively small set of products over which local managers

had more pricing power. Second, supermarkets in states that restricted the ways in which

scanners’ efficiency gains could be realized reduced prices less than supermarkets facing less

regulation. These two findings suggest that pass-through of efficiency gains from scanners

depended on stores’ flexibility with respect to pricing and other business practices. This,

in turn, implies that scanners could have had a bigger impact on prices if supermarkets’

28

practices had been more flexible, both with respect to state regulation and with respect to

firm-level control. In the taxonomy of Bloom, Garciano, Sadun, and Reenen (2011), scanners

are an information, not a communication, technology; their effectiveness depends on allowing

lower-level employees (store managers or even department managers) to make key decisions.

Finally, I find that later adopters contributed much less to price decreases than early

adopters, despite improvements in the technology. This finding suggests that, like other

process innovations, scanners were only as productivity-enhancing as the supermarkets that

adopted them. In particular, scanners were more efficient when complementary technologies

and processes were also adopted, and later adopters were either less able to make these

complementary changes or delayed these changes beyond the window of this sample.

Technology complements rather than substitutes for management practices, and efficient

practices have a cumulative effect on productivity (see, e.g., Aral, Brynjolfsson, and Wu,

2012). This paper contributes to the discussion on the importance of specific innovations

by showing that even the relatively “dumb” early scanners had a large effect on consumer

prices, but the heterogeneity of this effect across stores demonstrates the importance of

complementary processes. Process innovation can have large effects, but the gains from a

new process depend on the margins over which the adopting firm can make adjustments,

and on the speed of these adjustments.

29

References

Aaronson, D. (2001) “Price Pass-through and the Minimum Wage,” Review of Economicsand Statistics, 83(1), 158–169.

Anderson Economic Group (2010) “Michigan’s Item Pricing Law: The Price Tag for Retailersand Consumers,” Commissioned by Michigan Retailers Association.

Aral, S., E. Brynjolfsson, and L. Wu (2012) “Three-Way Complementarities: PerformancePay, Human Resource Analytics, and Information Technology,” Management Science,58(5), 913–931.

Barron, J. (1984) “Repricing Confuses Shoppers,” New York Times, February 25, 1984, 48.

Basker, E. (2005) “Selling a Cheaper Mousetrap: Wal-Mart’s Effect on Retail Prices,” Jour-nal of Urban Economics, 58(2), 203–229.

(2012) “Raising the Barcode Scanner: Technology and Productivity in the RetailSector,” American Economic Journal: Applied Economics, 4(3), 1–29.

Basker, E., and M. Noel (2009) “The Evolving Food Chain: Competitive Effects of Wal-Mart’s Entry into the Supermarket Industry,” Journal of Economics and ManagementStrategy, 18(4), 977–1009.

Beck, J., M. Grajek, and C. Wey (2011) “Estimating Level Effects in Diffusion of a NewTechnology: Barcode Scanning at the Checkout Counter,” Applied Economics, 43(14),1737–1748.

Bergen, M., D. Levy, S. Ray, P. H. Rubin, and B. Zeliger (2008) “When Little Things Meana Lot: On the Inefficiency of Item-Pricing Laws,” Journal of Law and Economics, 51(2),209–250.

Binkley, J. K., and J. M. Connor (1998) “Grocery Market Pricign and the New CompetitiveEnvironment,” Journal of Retailing, 74(2), 273–294.

Bloom, G. F. (1972) Productivity in the Food Industry. MIT Press, Cambridge, MA.

Bloom, N., L. Garciano, R. Sadun, and J. V. Reenen (2011) “The Distinct Effects of In-formation Technology and Communication Technology on Firm Organization,” StanfordUniversity Working Paper.

Brown, S. A. (1997) Revolution at the Checkout Counter: The Explosion of the Bar Code.Harvard University Press, Cambridge, MA.

Brynjolfsson, E., and L. M. Hitt (2003) “Computing Productivity: Firm-Level Evidence,”Review of Economics and Statistics, 85(4), 793–808.

Chevalier, J. A. (1995) “Do LBO Supermarkets Charge More? An Empirical Analysis of theEffects of LBOs on Supermarket Pricing,” Journal of Finance, 50(4), 1095–1112.

30

Chevalier, J. A., A. K. Kashyap, and P. E. Rossi (2003) “Why Don’t Prices Rise DuringPeriods of Peak Demand? Evidence from Scanner Data,” American Economic Review,93(1), 15–37.

Chevalier, J. A., and D. S. Scharfstein (1996) “Capital-Market Imperfections and Counter-cyclical Markups: Theory and Evidence,” American Economic Review, 86(4), 703–725.

Choi, C.-Y., and J. Wu (2012) “Understanding Retail Price Differences across U.S. Cities:A Tale of Two Frictions,” unpublished paper, University of Texas at Arlington.

Cortes, P. (2008) “The Effect of Low-Skilled Immigration on U.S. Prices: Evidence fromCPI Data,” Journal of Political Economy, 116(3), 381–422.

Courtemanche, C. J., and A. Carden (2012) “Competing with Costco and Sam’s Club: Ware-house Club Entry and Grocery Prices,” National Bureau of Economic Research WorkingPaper 17220.

Das, N., E. M. Falaris, and J. G. Mulligan (2009) “Vintage Effects and the Diffusion ofTime-Saving Technological Innovations: The Adoption of Optical Scanners by U.S. Su-permarkets,” B.E. Journal of Economic Analysis and Policy (Advances), 9(1), article 23.

Distribution Codes, Inc. (1974) “Directory of UGPCC Members,” Membership as of July31, 1974.

Doms, M. E., R. S. Jarmin, and S. D. Klimek (2004) “Information Technology Investment andFirm Performance in U.S. Retail Trade,” Economics of Innovation and New Technology,13(7), 595–613.

Einav, L., E. Leibtag, and A. Nevo (2010) “Recording Discrepancies in Nielsen HomescanData: Are They Present and Do They Matter?,” Quantitative Marketing and Economics,8(2), 207–239.

Food Fair, Inc. (1977) Annual Report. Philadelphia, PA.

Food Marketing Institute (various years) “Scanning Installation Up-Date,” Washington, DC.

Foster, L., J. Haltiwanger, and C. J. Krizan (2002) “The Link between Aggregate and Mi-cro Productivity Growth: Evidence from Retail Trade,” National Bureau of EconomicResearch Working Paper 9120.

(2006) “Market Selection, Reallocation and Restructuring in the U.S. Retail TradeSector in the 1990s,” Review of Economics and Statistics, 88(4), 748–758.

Frankel, D. M., and E. D. Gould (2001) “The Retail Price of Inequality,” Journal of UrbanEconomics, 49(2), 219–239.

The Great Atlantic and Pacific Tea Company (1976) Great Atlantic and Pacific Tea CompanyAnnual Report. Montvale, NJ.

31

Haberman, A. L. (2001) Twenty-Five Years behind Bars: The Proceedings of the Twenty-Fifth Anniversary of the U.P.C. at the Smithsonian Institution, September 30, 1999. Har-vard University Press, Cambridge, MA.

Haskel, J., and R. Sadun (2009) “Entry, Exit and Labour Productivity in UK Retailing:Evidence from Micro Data,” in Producer Dynamics: New Evidence from Micro Data, ed.by J. B. Jensen, T. Dunne, and M. J. Roberts. University of Chicago Press.

Hausman, J., and E. Leibtag (2007) “Consumer Benefits from Increased Competition inShopping Outlets: Measuring the Effect of Wal-Mart,” Journal of Applied Econometrics,22(7), 1157–1177.

Hoch, S. J., A. L. Montgomery, and Y.-H. Park (2000) “Long-Term Growth Trends in PrivateLabel Market Shares,” Wharton Marketing Department Working Paper 00-010.

Holmes, T. (2001) “Bar Codes Lead to Frequent Deliveries and Superstores,” RAND Journalof Economics, 32(4), 708–725.

Hurst, J. B. (2007) “Time to Take the Next Step on Item Pricing Reform,” The RetailReview, September/October, 2.

International Business Machines Corporation (1980) Annual Report. Armonk, NY.

Krishnamurthi, L., and S. P. Raj (1991) “An Empirical Analysis of the Relationship BetweenBrand Loyalty and Consumer Price Elasticity,” Marketing Science, 10(2), 172–183.

The Kroger Co. (various years) Kroger Co. Annual Report. Cincinnati, OH.

Lach, S. (2007) “Immigration and Prices,” Journal of Political Economy, 115(4), 548–587.

Lagakos, D. (2013) “Explaining Cross-Country Productivity Differences in Retail Trade,”unpublished paper, Arizona State University.

Levin, S. G., S. L. Levin, and J. B. Meisel (1985) “Intermarket Differences in the EarlyDiffusion of an Innovation,” Southern Economic Journal, 51(3), 672–80.

(1987) “A Dynamic Analysis of the Adoption of a New Technology: The Case ofOptical Scanners,” Review of Economics and Statistics, 69(1), 12–17.

(1992) “Market Structure, Uncertainty, and Intrafirm Diffusion: The case of OpticalScanners in Grocery Stores,” Review of Economics and Statistics, 74(2), 345–350.

Levy, D. (2008) “Item Pricing Laws: Review of World Status, Emerging Trends, and Esti-mates of the Law’s Impact on Consumer Prices,” Report for the Department of Researchand Economics, Israeli Ministry of Industry, Trade, and Labor (in Hebrew).

Levy, D., M. Bergen, S. Dutta, and R. Venable (1997) “The Magnitude of Menu Costs:Direct Evidence from Large U.S. Supermarket Chains,” Quarterly Journal of Economics,12(3), 791–825.

32

Marr, K. (1996) “Will the Real Price Please Stand Up!,” unpublished manuscript.

Matsa, D. A. (2011) “Competition and Product Quality in the Supermarket Industry,”Quarterly Journal of Economics, 126(3), 1539–1591.

McEnally, M. R., and J. M. Hawes (1984) “The Market for Generic Brand Grocery Products:A Review and Extension,” Journal of Marketing, 48(1), 75–83.

McGrayne, S. (1986) “Supermarkets Eye Price-Code Scanners,” Crains Detroit Business,February 3, 1986, 3.

Nakamura, L. I. (1999) “The Measurement of Retail Output and the Retail Revolution,”Canadian Journal of Economics, 32(2), 408–425.

NCR (various years) Annual Report. Dayton, OH.

Parsley, D. C., and S.-J. Wei (1996) “Convergence to the Law of One Price without TradeBarriers or Currency Fluctuations,” Quarterly Journal of Economics, 111(4), 1211–1236.

Progressive Grocer (1976) “‘Electronic’ Produce Pays Off Big for Family Super,” ProgressiveGrocer, 55(3), 40–43.

Richburg, K. B. (1981) “15 Stores Stop Marking Price on Individual Items,” WashingtonPost, July 3, 1981, B2.

Safeway Inc. (various years) Safeway Inc. Annual Report. Pleasanton, CA.

Selmeier, B. (2008) Spreading the Barcode: Personal Memories of Bill Selmeier. Lulu.

Sieling, M., B. Friedman, and M. Dumas (2001) “Labor Productivity in the Retail TradeIndustry, 1987-99,” Monthly Labor Review, December, 3–14.

Singh, V. P., K. T. Hansen, and R. C. Blattberg (2006) “Market Entry and ConsumerBehavior: An Investigation of a Wal-Mart Supercenter,” Marketing Science, 25(5), 457–476.

Supermarkets General Corporation (1982) Annual Report. Woodbridge, NJ.

Syverson, C. (2011) “What Determines Productivity?,” Journal of Economic Literature,49(2), 326–365.

Ward, M. B., J. P. Shimshack, J. M. Perloff, and J. M. Harris (2002) “Effects of the Private-Label Invasion in Food Industries,” American Journal of Agricultural Economics, 84(4),961–973.

Williams, W. (1984) “The New Supermarket Takes Shape,” New York Times, September 2,1984.

Wilson, T. (1976) “Where Do We Stand on UPC Implementation,” Journal of Food Distri-bution Research, February, 96–102.

33

Winn-Dixie Stores, Inc. (1979) Annual Report. Jacksonville, FL.

Zimmerman, G. E. (1999) “The Effects of Network Externalities on the Adoption of the Uni-versal Product Code: An Empirical Investigation of Grocery Retailers and Manufacturers,1970–1980,” Harvard University undergraduate honors thesis.

34

Figure 1. Mean and Inter-Quartile Range of Six Real Price Series

Figure 2. U.S. and Sample Scanner Installations, 1972–1985q1

35

Figure 3. Estimated SR Effects of 10% Scanner Penetration, by Product

36

Table 1. ACCRA Retail Price Summary Statistics

Average StandardProduct Description Obs Price Deviation

Produce:Bananas 1 lb. 7,058 0.91 0.17Lettuce Head iceberg lettuce, approx. 11/2 lbs. 7,059 1.75 0.45Potatoes 10 lbs., red or white 7,029 4.79 1.61

Dairy:Eggs 1 dozen, grade A, large 7,058 2.58 0.62Milk 1/2 gallon whole milk, carton, fresh 7,058 3.50 0.44Margarine 1 lb., cheapest 7,058 1.44 0.42

Meat:Beef Ground beef, 1 lb. 7,059 3.71 0.78Bacon 1 lb., cheapest 7,058 4.27 1.70Chicken Frying chicken, grade A, whole, per lb. 7,059 2.10 0.44

Canned:Peaches #21/2 (approx 29 oz.) can halved peaches, cheapest 7,059 2.30 0.26Baby food 4.5 oz. jar Gerber’s strained vegetables 6,256 0.64 0.80Tomatoes #303 can (15–17 oz.), cheapest 7,059 1.25 0.19

Misc.:Orange Juice 6 oz. can, frozen, cheapest 7,055 1.17 0.18Bread 20 oz. white bread, cheapest 7,058 1.55 0.32Shortening 3 lbs. can all vegetable shortening, Crisco brand 7,058 6.19 1.50

Non-Grocery:

Gasoline1 gallon, unleaded regular grade, including

7,057 2.65 0.52taxes, national brand, self-service if available

Cigarettes Carton, Winston, king size 7,058 16.85 2.26Whiskey Seagram’s 7 Crown, 750 ml. 6,989 19.55 4.23All prices are in December 2012 dollarsMost missing observations are due to unreadable hard-copya Gerber’s baby food is available only to 1983q4

37

Table 2. ACCRA City Summary Statistics

StandardVariable Average Deviation Min Max

Scanners 3.1 10.1 0 225Food-selling establishments (SIC 54) 161.6 277.5 5 3,764Scanners/SIC54 2.5% 5.3% 0 52.3%

Grocery establishments (SIC 541) with payrolla 127.0 163.0 15 1,362Scanners/SIC541 3.9% 7.9% 0 85.2%a SIC 541 with payroll only available for a 140 cities

Table 3. BLS Metropolitan Areas, 1972–1984

AverageMetro Area States Inflationa

Anchorage AK 0.53%Atlanta GA 0.55%Baltimore MD 0.52%Boston-Lawrence-Salem MA-NH 0.48%Buffalo-Niagara Falls NY 0.51%Chicago-Gary-Lake County IL-IN-WI 0.49%Cincinnati-Hamilton OH-KY-IN 0.57%Cleveland-Akron-Lorain OH 0.53%Dallas-Fort Worth TX 0.54%Denver-Boulder CO 0.50%Detroit-Ann Arbor MI 0.51%Honolulu HI 0.60%Houston-Galveston-Brazoria TX 0.56%Kansas City MO-KS 0.51%Los Angeles-Anaheim-Riverside CA 0.54%Miami-Ft. Lauderdaleb FL 0.54%Milwaukee WI 0.51%Minneapolis-St. Paul MN-WI 0.52%New York-Northern New Jersey-Long Island NY-NJ-CT-PA 0.53%Philadelphia-Wilmington-Trenton PA-NJ-DE 0.51%Pittsburgh-Beaver Valley PA 0.48%Portland-Vancouver OR-WA 0.52%San Diego CA 0.54%San Francisco-Oakland-San Jose CA 0.55%Seattle-Tacoma WA 0.50%St. Louis-East St. Louis MO-IL 0.55%Washington DC-MD-VA 0.55%aAverage monthly rate across all categoriesbMiami-Ft. Lauderdale starts 11/1977

38

Table 4. BLS CPI Series, 1972–1984

AverageSeries Description Inflationa

Food Food at home 0.54%Produce Fruits and vegetables 0.57%Dairy Dairy products 0.48%Meat Meats, poultry, fish, and eggsb 0.33%Bakery Cereals and bakery products 0.61%Other Other food at homeb 0.60%aAverage monthly rate across all MSAsbStarts in 1976

Table 5. Retail Prices as a Function of Scanners’ Share of SIC 54

All Produce Dairy Meat Canned Misc.

ScanFrac -0.1429*** -0.2811*** -0.0726 -0.2069*** -0.0236 -0.1208*(0.0487) (0.0817) (0.0540) (0.0630) (0.0448) (0.0680)

Products 15 3 3 3 3 3Product×City FE X X X X X XProduct×Time FE X X X X X XObservations 105,041 21,146 21,174 21,176 20,374 21,171SR Effect -0.0143 -0.0281 -0.0073 -0.0207 -0.0024 -0.0121Each column represents a separate regressionRobust standard errors in parentheses, clustered by city* p<10%; ** p<5%; *** p<1%

The estimated SR effect is 0.10 · θ, where θ is the coefficient on ScanFrac

Table 6. Retail Prices as a Function of Scanners’ Share of SIC 54: IPL vs. non-IPL

All Produce Dairy Meat Canned Misc.

ScanFrac×IIPL=0 -0.1580*** -0.2650*** -0.1110** -0.2156*** -0.0462 -0.1433**(0.0497) (0.0846) (0.0529) (0.0639) (0.0460) (0.0711)

ScanFrac×IIPL=1 0.0490 -0.4637*** 0.4022*** -0.0885 0.2580*** 0.1594(0.0870) (0.1474) (0.1150) (0.1141) (0.0838) (0.1242)

Products 15 3 3 3 3 3Product×City FE X X X X X XProduct×Time FE X X X X X XObservations 105,041 21,146 21,174 21,176 20,374 21,171F testa 5.9267 1.8938 21.1702 1.4439 14.7518 5.4807p value 0.0159 0.1704 0.0000 0.2311 0.0002 0.0203Each column represents a separate regressionRobust standard errors in parentheses, clustered by citya Test for equal coefficients* p<10%; ** p<5%; *** p<1%

39

Table 7. Retail Prices as a Function of Scanners’ Share of SIC 54 by Vintage

All Produce Dairy Meat Canned Misc.

ScanFracV.1&2 -0.2188*** -0.3718*** -0.1693* -0.2361*** -0.1483* -0.1654*(0.0729) (0.1288) (0.0886) (0.0881) (0.0765) (0.0977)

ScanFracV.3 -0.0854 -0.2126** -0.0002 -0.1838** 0.0728 -0.0870(0.0554) (0.0980) (0.0571) (0.0729) (0.0598) (0.0834)

Products 15 3 3 3 3 3Product×City FE X X X X X XProduct×Time FE X X X X X XObservations 105,041 21,146 21,174 21,176 20,374 21,171F testa 2.7497 1.1650 3.2686 0.2693 5.6213 0.4256p value 0.0990 0.2818 0.0722 0.6044 0.0188 0.5150Each column represents a separate regressionRobust standard errors in parentheses, clustered by citya Test for equal coefficients* p<10%; ** p<5%; *** p<1%

Table 8. Retail Prices as a Function of Scanners’ Share of SIC 541 with Payroll

All Produce Dairy Meat Canned Misc.

ScanFrac∗ -0.1356*** -0.2630*** -0.0843* -0.1813*** -0.0306 -0.1098**(0.0407) (0.0641) (0.0446) (0.0512) (0.0419) (0.0537)

Products 15 3 3 3 3 3Product×City FE X X X X X XProduct×Time FE X X X X X XObservations 78,064 15,707 15,735 15,737 15,153 15,732SR Effect -0.0217 -0.0421 -0.0135 -0.0290 -0.0049 -0.0176Each column represents a separate regressionRobust standard errors in parentheses, clustered by city* p<10%; ** p<5%; *** p<1%

The estimated SR effect is 0.16 · θ, where θ is the coefficient on ScanFrac∗

Table 9. CPI Levels as a Function of Scanners’ Share of SIC 541 with Payroll

Food Produce Dairy Meat Bakery Other

ScanFrac∗ -0.0719* -0.1980 -0.0174 -0.1836** -0.0106 -0.1112*(0.0355) (0.1258) (0.0514) (0.0729) (0.0636) (0.0592)

MSA FE X X X X X XTime FE X X X X X XObservations 4,102 4,101 4,101 2,894 4,101 2,894Each column represents a separate regressionRobust standard errors in parentheses, clustered by MSA* p<10%; ** p<5%; *** p<1%

40

Table 10. Non-Grocery Retail Prices as a Function of Scanners’ Share of SIC 54

All Gasoline Cigarettes Whiskey

ScanFrac -0.0146 -0.0246 0.0357 -0.0552(0.0257) (0.0331) (0.0503) (0.0695)

Products 3 1 1 1Product×City FE X X X XProduct×Time FE X X X XObservations 21,104 7,058 6,989 7,057Each column represents a separate regressionRobust standard errors in parentheses, clustered by city* p<10%; ** p<5%; *** p<1%

Table 11. Retail Prices as a Function of Current and Future Scanner Share of SIC 54

All Produce Dairy Meat Canned Misc.

ScanFraci,t+1 0.0348 -0.1301 0.0502 0.0135 0.0911 0.1771(0.0765) (0.1425) (0.0808) (0.1204) (0.1014) (0.1216)

ScanFrac -0.1782** -0.1520 -0.1264 -0.2003 -0.1209 -0.2844**(0.0748) (0.1352) (0.0809) (0.1265) (0.1026) (0.1230)

Products 15 3 3 3 3 3Product×City FE X X X X X XProduct×Time FE X X X X X XObservations 105,041 21,146 21,174 21,176 20,374 21,171Each column represents a separate regressionRobust standard errors in parentheses, clustered by city* p<10%; ** p<5%; *** p<1%

41