Retail Prices in a City∗

Alon Eizenberg†1, Saul Lach‡2, and Merav Oren-Yiftach§3

1The Hebrew University of Jerusalem & CEPR2The Hebrew University of Jerusalem & CEPR

3Israel Central Bureau of Statistics

June 9, 2020

Abstract

This study examines grocery price differentials across neighborhoods in a large metropolitanarea (the city of Jerusalem, Israel). Important variation in access to affordable grocery shoppingis documented using CPI data on prices, and neighborhood-level credit card expenditure data.Residents of peripheral, non-affluent neighborhoods are charged some of the highest prices in thecity, and yet display a low tendency to shop outside their neighborhood. In contrast, residents ofaffluent, centrally-located neighborhoods often benefit from lower grocery prices charged in theirown neighborhood, while also displaying a high propensity to shop at the hard-discount grocerslocated in the city’s commercial districts. The role of spatial frictions in shaping these patternsis studied within a structural model where households determine their shopping destination andretailers choose prices. The estimated model implies strong spatial segmentation in households’demand. Counterfactual analyses reveal that alleviating spatial frictions results in considerablebenefits to the average resident of the peripheral neighborhoods. At the same time, it barelyaffects the equilibrium prices charged across the city, and so it does little to benefit householdswith limited mobility (e.g., the elderly).

∗We thank Eyal Meharian and Irit Mishali for their invaluable help with collecting the price data and withthe provision of the geographic (distance) data. We also wish to thank a credit card company for graciouslyproviding the expenditure data. We are also grateful to Daniel Felsenstein for providing the housing pricedata, and to Elka Gotfryd for mapping zipcodes into statistical subquarters. We are grateful to the editor andanonymous referees for comments and suggestions which greatly improved the paper. We thank Steve Berry,Pierre Dubois, Phil Haile, JF Houde, Gaston Illanes, Volker Nocke, Kathleen Nosal, Mark Rysman, Katja Seim,Avi Simhon, Konrad Stahl, Yuya Takashi, Ali Yurukoglu and Christine Zulehner for helpful comments, as wellas seminar participants at Carlos III, CEMFI, DIW Berlin, Frankfurt, Harvard, Johns Hopkins, Penn State,Universidad de Vigo, UVA, Yale and Wharton, and participants at the Israeli IO day (2014), EARIE (2014),the Economic Workshop at IDC (2015), UTDT Conference (2016), CEPR-JIE IO Conference (2017), and IIOC(2017). This project was supported by the Israeli Science Foundation (ISF) grant 858/11, by the Wolfson FamilyCharitable Trust, and by the Maurice Falk Institue for Economic Research in Israel.†alon.eizenberg@mail.huji.ac.il‡saul.lach@mail.huji.ac.il§meravo@cbs.gov.il

1 Introduction

In January 2014, residents of Qiryat HaYovel, a residential neighborhood in the city of Jerusalem,

Israel, initiated a consumer boycott against a neighborhood supermarket. They claimed that

prices at this supermarket were much higher than those charged at branches of the same chain

operating elsewhere in the city, and that such alternative shopping destinations were not readily

accessible to them: “Young families will not travel to Talpiot or Givat Shaul (two commercial

districts with hard discount supermarkets) to shop and, instead, shop in the neighborhood

for lack of time.”1 Senior residents of the neighborhood were mentioned as another affected

demographic group. The boycott ended after the chain agreed to lower the cost of a basket of

goods by 14 percent, according to the organizers. In June 2017, a similar consumer boycott in

the Jerusalem neighborhood of Gilo was reported to have achieved its goals: the management

agreed to equate prices there to those charged by the same chain at one of the city’s large

commercial districts.2

The “boycotting” neighborhoods share important characteristics: both are non-affluent

neighborhoods located in the periphery of the city, at considerable distance from the main

shopping districts. Access to affordable grocery shopping may therefore depend on both so-

cioeconomic factors and geographic location. In this paper we systematically explore these

relationships. Using price data collected in Jerusalem by the Central Bureau of Statistics

(CBS), we document considerable variation in grocery prices charged across neighborhoods.

We also explore households’ shopping patterns via data on aggregate grocery expenditure flows

between neighborhoods, obtained from a credit card company.

The combined message from these data is striking: residents of non-affluent, peripheral

neighborhoods are charged very high prices in their own neighborhood but, despite this fact,

have a high tendency to shop in it. In contrast, residents of more affluent, centrally-located

neighborhoods are often charged lower prices in their neighborhood, and also display a higher

tendency to shop at even cheaper locations outside their neighborhood. Both empirical facts

are consistent with activists’ complaints regarding the lack of access to affordable shopping

locations. In equilibrium, this lack of access enhances the market power enjoyed by retailers

operating in the peripheral neighborhoods.3

We explore these equilibrium forces by estimating an empirical model of demand and supply

for groceries. Our demand model quantifies households’ price-distance trade-off which is sub-

stantial: “pushing” a destination neighborhood 1 km further away from an origin neighborhood

reduces demand from that origin by about 35 percent. The supply model describes retailers’

1“Qiyat HaYovel: the residents’ battle against ‘My Shufersal’,” Ynet (an Israeli news outlet), January 2014.2The organizers said that the chain insisted on maintaining pricing flexibility on 20 “non-essential” products.

“Shufersal Deal Gilo announced it will set the same prices as charged at the Talpiot branch,” Kol Hair (aJerusalem local newspaper), June 2017.

3Such market power could attract additional supermarket entry into these neighborhoods. However, bar-riers to supermarket entry in residential neighborhoods are substantial owing to space constraints and zoningrestrictions. For tractability reasons, in this paper we treat the entry decisions of supermarkets as given.

1

pricing decisions given the estimated demand elasticities.

We use the model in counterfactual analysis to tease out the impact of spatial frictions. We

investigate the effects of reducing these frictions by changing parameters that capture house-

holds’ implicit cost of travel. We find that the price level charged in residential neighborhoods is

only mildly reduced, and, in certain peripheral neighborhoods, prices even go up slightly. This

seemingly-surprising result captures the effect of two conflicting forces: increased competition

exerts a downward pressure on prices but, at the same time, the average demand elasticity

faced by residential neighborhood retailers becomes lower, prompting them to raise prices.

Whereas prices barely respond, consumer behavior does change when spatial frictions are

alleviated: a much larger fraction of residents now travels to the hard discount supermarkets

in the commercial districts. As a consequence, the average price paid by residents of a given

neighborhood declines significantly, and this effect is particularly strong in the peripheral,

non-affluent neighborhoods. An important lesson from this finding is that reductions in spatial

frictions can have important policy effects that would be entirely missed by a statistical analysis

of prices alone. This motivates our joint analysis of both prices and shopping patterns.

Another lesson from this analysis is that, while the average resident of the peripheral, non-

affluent neighborhoods gains substantially when spatial frictions are alleviated, residents with

reduced mobility (e.g., the elderly) gain little or not at all, as equilibrium prices charged in

those neighborhoods remain high. Relief to such residents will likely require targeted policies.

Our analysis reveals the connections between spatial frictions and the distribution of grocery

prices, and the manner by which different groups of households may be affected by potential

changes to such frictions. While our counterfactual exercises are conceptual and not intended

to simulate concrete policies, they do have some policy relevance. In the case of Jerusalem,

the city plans to improve access to its main shopping district both via the extension of the

light rail system and via improvements to its internal organization, actions that mimic our

counterfactuals.4

We next explain how we differ from previous analyses of spatial price equilibria within urban

settings. Following this literature review, the paper proceeds as follows: Section 2 presents our

data, Section 3 presents the demand model and its estimation, and Section 4 describes our

pricing model. Section 5 presents our counterfactual experiments, and Section 6 concludes.

Related literature. Retail price differentials across neighborhoods have attracted consid-

erable attention in the economic literature. Such differentials suggest that standard measures

of inequality, based on nominal wages, may be biased (as in Moretti 2013). A vast literature,

starting with Caplovitz (1963), has attempted to measure such price differentials to understand

whether “the poor pay more,” producing mixed empirical findings.5 This literature focuses on

4“The plan: the Talpiot industrial zone to undergo a revolution in the next decade,” Kol Hair, April 2016.5MacDonald and Nelson (1991), for example, compared the price of a fixed basket of goods across 322

supermarkets in 10 metropolitan areas in the US, revealing that prices in suburban locations were about 4percent lower than in central city stores where poorer population lived. Chung and Myers (1999) similarlyreport that the price of a weekly home food plan was higher in poorer neighborhoods of the Twin Cities

2

statistical comparisons of retail prices across affluent and non-affluent neighborhoods. With

relatively few exceptions, these studies have abstracted from the issue of cross-neighborhood

shopping (i.e., shopping outside one’s neighborhood of residence).6 We differ from this litera-

ture in two ways: we combine the typical neighborhood-level price data with less typical data

on cross-neighborhood shopping flows, and move beyond statistical comparisons by presenting

a structural model of the equilibrium forces driving the observed price differentials.

A vast literature on spatial frictions includes classic theoretical contributions by Hotelling

(1929) and Salop (1979). Several recent empirical papers have taken a structural approach to

study spatial competition in various industries, including Adams and Williams (2019), Miller

and Osborne (2012), Thomadsen (2005), Davis (2006), McManus (2007), Houde (2012) and

Davis, Dingel, Monras and Morales (2019). Substantial empirical work has considered spatial

competition among supermarkets (for example, Chintagunta, Dube, and Singh 2003, and Smith

2004). Our work shares several features with Dubois and Jodar-Rosell (2010) who study super-

market competition: we also estimate a discrete-continuous demand model, use a supply-side

model to identify heterogeneous marginal costs, and consider a counterfactual analysis in which

travel costs are reduced (see also Figurelli 2013 and Ellickson, Grieco, and Khvastunov 2016).

Although we share with the Industrial Organization (IO) literature the structural approach

to estimate demand and supply primitives, our work is motivated by a perennial question in

the urban economics literature. Our focus on the relationship between prices, location and

consumer flows motivates our use of data sources different from the typical scanner data used

in IO papers. Scanner data are ideal for uncovering rich preference structures, but they may

be less useful for uncovering shopping patterns within the city. Instead we use a price index

for a basket of grocery goods derived from the official statistical agency’s methodology that

is comparable across space and time, as well as credit card data that provide a systematic

description of shopping flows across all neighborhoods. While it is possible to construct such

neighborhood-level price indices and consumer flows from scanner data, it is not obvious that

doing so will provide sufficient coverage in the context of our research question.7 Overall,

we view our approach as complementary to the established use of scanner data in studying

supermarket demand.

metropolitan area. Recent work, in contrast, reports that prices in richer zip codes (Hayes, 2000) or prices paidby high income households (Aguiar and Hurst, 2007) are significantly higher.

6Kurtzon and McClelland (2010) study a Bureau of Labor Statistics survey in which respondents report theirshopping destinations. They find that the “poor pay neither more nor less than the rich at the stores they shopat.” See also Aguiar and Hurst (2007) and Griffith et al. (2009) for analyses of survey data where recordedprices correspond to prices actually paid by households.

7Even if the sample of households in the scanner data is random and representative of residents in eachneighborhood, it need not adequately cover all origin-destination neighborhood pairs characterizing the shoppingdecisions. Our credit card data also suffer from selectivity bias, and we address this issue econometrically.

3

2 Data

We begin by describing Jerusalem’s urban structure and its notable partition into distinct neigh-

borhoods. Additional subsections describe the prices collected at retail locations throughout

the city, and the data on consumer transaction flows across neighborhoods.

2.1 Jerusalem’s urban structure: neighborhoods

Our analysis covers 46 neighborhoods in Western Jerusalem (see Online Appendix C for def-

initions). These are predominantly Jewish neighborhoods. The eastern part of the city has

predominantly Arab neighborhoods which we do not include in our study because of significant

differences in the basket of groceries purchased and in the extent of credit card usage across

these populations. Moreover, residents of Western Jerusalem do not typically perform their

weekly grocery shopping in Eastern Jerusalem and vice-versa.

Figure 1 displays Jerusalem’s neighborhoods, highlighting those covered by our study in

color. Neighborhoods developed historically along the roads radiating from the “Old City,”

as is typical in many ancient cities. Jerusalem’s hilly topography resulted in geographically

separated neighborhoods such that moving between them typically requires some mode of

transportation. Most neighborhoods have a small commercial center with a small grocery store

and other retail services while many, but not all, have one or two supermarkets. Hard discount

(HD) supermarkets are located in well-defined commercial districts. Jerusalem does not have

an important suburban ring surrounding it.

At the neighborhood level, we observe demographic variables (from the 2008 Israel Census

of Population) that are likely to shift price and travel sensitivities: the fraction of the neigh-

borhood’s households that own a car, the fraction of residents above the age of 15 who drive

to work, and the fraction of senior residents. We use the average price of housing per square

meter in 2007-2008, obtained from the Tax Authority’s records of real estate transactions, as a

proxy for the neighborhood’s wealth. Table 1 reports descriptive statistics.

There are sharp socioeconomic differences across neighborhoods. For example, housing is

more than twice as expensive in the central, affluent neighborhood of Rehavya than in the

peripheral, non-affluent neighborhood of Neve Yaaqov. In our model, this variation will help

identify price and distance sensitivities.

Spatial frictions are captured via a 46-by-46 matrix of distances between each pair of neigh-

borhoods.8 Table 2 reports statistics regarding the distance to the city center and to the city’s

two prominent commercial districts (Talpiot and Givat Shaul, hereafter referred as CD1 and

8In the online Appendix C we explain that each neighborhood is comprised of several “statistical areas.”The CBS provided us with the shortest road distance between the centroids of each pair of such areas. Wethen compute the distance djn between neighborhoods j and n as an average of the distances between eachpair of statistical areas belonging to these neighborhoods. As some neighborhoods are quite large, we defineneighborhood j’s “own distance” djj as the mean distance between the centroids of each pair of statistical areasincluded in it.

4

CD2, respectively). It also reports statistics on the average distance to all the other neighbor-

hoods, a rough measure of how peripheral the neighborhood is. Table 2 indicates considerable

variation: the maximum distances to the city center and to the commercial districts are about

twice as large as the corresponding mean (or median) distance.

To facilitate the study of our research question we next identify several neighborhoods of

interest that differ considerably in terms of their observed characteristics and location within

the city. Three neighborhoods — Neve Yaaqov, Givat Shapira, and Qiryat HaYovel South — are

both Non-Affluent and Peripherally located. We shall hereafter refer to them as NAP1, NAP2,

and NAP3, respectively. Appendix Table C2 shows that housing prices in those neighborhoods

are 9.5, 10.7, and 11.5 NIS, i.e., below the mean (median) of 13.4 (13.3) reported in Table 1.

These neighborhoods’ peripheral location is clearly indicated in Figure 1.

We also identify three other neighborhoods as Affluent and Centrally located: Rehavya,

Qiryat Moshe - Bet Hakerem, and Baqa-Abu Tor-Yemin Moshe, denoted by AC1, AC2 and AC3,

respectively. In these AC neighborhoods housing prices are 21.1, 15.8 and 15 NIS (Appendix

Table C2), well above the mean price. Figure 1 shows that these neighborhoods are within

close proximity to the city center, as well as to the CD1 and CD2 commercial districts.

In short, Jerusalem shares many of the characteristics of other large metropolitan areas: it

features well-defined commercial districts, affluent and less affluent neighborhoods, and central

and peripheral locations. It is therefore a useful laboratory to study the role of spatial frictions

in generating price differentials across neighborhoods.

2.2 Price data

The price data were collected by CBS personnel as part of their monthly computation of

the Consumer Price Index (CPI), but the sample used in this research includes additional

supermarkets, beyond those normally used in the CPI sample. We focus on 27 popular, everyday

products consumed by most households. Each selected product is associated with a unique

universal product code (UPC) and is therefore identical across sampled stores (e.g., the same

brand, size, packaging, etc.).9 Price observations were collected in three periods: in September

and November 2007, and in November 2008. CBS personnel sampled 60 distinct stores in

Jerusalem: about 55 percent of them were supermarkets, 20 percent were open market stalls and

15 percent were grocery stores. The sampled stores are present in 26 of our 46 neighborhoods.

While this may appear as a major omission, we note that in the remaining 20 neighborhoods

there are no important supermarkets and, typically, they only have a small grocery store.10

The list of products, their mean price and coefficients of variation are displayed in Online

Appendix D. Fruits and vegetables usually exhibit higher price dispersion than other foodstuff.

9Even among fruits and vegetables there are no noticeable quality differences across stores because the CBScollects prices of produce of a specific quality grade.

10In our econometric demand model, we address the presence of neighborhoods without sampled prices in aninternally-consistent fashion by including such shopping destinations in the households’ outside option.

5

One possible explanation for this higher variance is their perishable nature: unsold stocks trigger

price reductions, thereby generating a higher variance. We further note that an alternative

composite good that excludes fruits and vegetables (as opposed to the one we use, defined

below, that includes all products) displays higher price dispersion across neighborhoods. The

dispersion in the prices of fruits and vegetables, while important, is therefore not the primary

driver of the price variation that we study in the paper.

A composite good. We aggregate individual product items to a composite good whose

price is measured at the neighborhood level. Our focus on a composite good is in line with the

relevant urban economics literature (e.g., MacDonald and Nelson 1991), and makes particular

sense in our application because we observe neighborhood-level expenditure flows.

Residential neighborhoods tend to be served by smaller, more expensive store formats,

whereas commercial neighborhoods have larger HD stores. Most price variation is, therefore,

between rather than within neighborhoods. Indeed, using the neighborhoods with at least

two stores, we computed the between and within variance of price for each item and period

separately. In 86 percent of the cases the between-neighborhood variance of prices is larger

than the within variance and the median ratio of between to within variance is 3.2.

We define the price of the composite good charged in a given neighborhood as a weighted

average of individual-item prices using the CPI expenditure weights. Letting ωi be the weight

of product i, i = 1, . . . , 27, Ωnt be the set of products observed in neighborhood n at time t,

and pnit the average price of product i in neighborhood n in period t (over all stores selling the

product in the neighborhood), the price of a single unit of the composite good is

(1) pnt =∑i∈Ωnt

(ωi∑

i∈Ωntωi

)pnit.

Missing price observations are typical of studies that construct indices from prices collected

by official statistical agencies. Indeed, not all 60 stores are surveyed in each of the three

periods, and not all products are surveyed in each store-period.11 Statistics regarding the

number of sampled stores and the prevalence of missing prices are presented in Table 3 (see

Online Appendix Table D3 for neighborhood-specific values). On average, the 26 neighborhoods

have two sampled stores, including one supermarket. CD1, the main commercial district, has

5 hard discount supermarkets. The average neighborhood has non-missing price data for 17-18

products, and the typical neighborhood (see bottom panel) has observed prices for most of the

27 products.

Our goal is to define a composite good that would be as homogeneous as possible without

reducing the sample size too much. We therefore pursue a leading specification that computes

11While our 27 items are popular products that should be available in all stores, recall that a product isdefined by its unique UPC. Some stores may carry a different version of what is essentially the same product(e.g., differing in packaging), generating a missing price observation.

6

the index in 15 neighborhoods (including four commercial districts) where at least 21 of the 27

items have a non-missing price observation. We treat prices in the remaining neighborhoods as

unobserved, a feature that will be consistent with our econometric model.12

One concern is that changes over time in the identity of the products in the composite

good can generate spurious price variation over time. Note, however, that in 8 out these 15

neighborhoods, we observe at least 26 of the 27 products in all three periods. In fact, most of

the products appear in all three periods in most neighborhoods. Another concern is that price

differences across neighborhoods reflect differences in the components of the composite good.

There is, however, considerable overlap in the basket of goods across neighborhoods.13

Nonetheless, to ensure that our results are not driven by spurious variation due to missing

data we projected, for each product separately, the observed prices on a large set of demographic

variables and used the estimated coefficients to impute prices in the neighborhoods with missing

prices. As reported below, the demand estimates using the imputed prices are qualitatively the

same as those in our leading specification. Thus, the limited changes in the composition of the

basket of goods over time or across neighborhoods do not appear to be driving our results.

Beyond this imputation exercise, we checked the robustness of our results to alternative

methods of computing the price of the composite good. We used a threshold lower than 21

products, we used only fruits and vegetables, and restricted the sample to supermarkets only.

The last two, in particular, ameliorate considerably the missing price problem. Reassuringly,

the estimated demand patterns remain qualitatively the same.

Using identical CPI weights for different households is commonplace in the literature, but

does not allow tastes to vary with income (Handbury 2013). The uniform weights result in

a well-defined single price at each location charged to residents of all origins. This makes

our counterfactual analyses more transparent. Nonetheless, we also computed a price index

using CPI weights that vary by socioeconomic standing, provided by the CBS. We thus assign

differential weights to different origin neighborhoods. This alternative price index has a simple

correlation of 0.85 with our index in equation (1) and, not surprisingly, delivers similar demand

estimates.14

Price differentials. Table 4 provides statistics for the price of the composite good. The

time variation in our sampled prices appears to be in line with the CPI inflation rate.15 Prices

vary significantly across neighborhoods within both commercial and residential districts, with

12The resulting subsample keeps essentially the same distribution of store formats as the 26 neighborhoodsample (57 percent supermarkets, 21 percent market stalls and 12 percent grocery stores).

13See Online Appendix Tables D3-D5.14All robustness checks are reported in online Appendix A. In our demand model we partially compensate

for the uniform weights by allowing households to derive utility from unobserved aspects of the shoppingdestinations captured by fixed effects, and by interacting those with the origin’s housing prices. We thereforeallow the variety of additional products (i.e., beyond our 27) to be valued differently by households of differentincome levels.

15The composite good’s price increased by 10 percent between November 2007 and November 2008. Forcomparison, the CPI inflation for food between December 2007 and December 2008 was 8.3 percent.

7

the maximum price being about 16-29 percent above the minimum price. The quantitative

importance of cross-neighborhood price variation is manifested in the (gross) savings generated

by shopping at the cheapest location in the city. Specifically, Panel C of Table 4 shows the

distribution of these savings, 100× (pjt −Minnpnt)/pjt for each of the 15 neighborhoods with

valid prices over all three periods; mean savings are 13 percent.

Another message of Table 4 is that prices in the commercial districts are in general lower

than in most residential neighborhoods; the mean price of the composite good is between 6-7

percent higher in the residential neighborhoods. Variation between residential neighborhoods is

also considerable: panel B shows that prices in our three NAP (non-affluent peripheral) neigh-

borhoods are typically ranked above prices in the three AC (affluent, central) neighborhoods.16

This observation is central to our research question and so we explore it using two additional

figures.

Figure 1 displays the composite good prices in November 2008. It corroborates the observa-

tion that some of the highest prices in the city are charged by retailers located in the peripheral,

non-affluent neighborhoods. Neighborhoods such as AC2 or AC3 are much more affluent, yet

retail prices charged there are lower than those charged at NAP1-NAP3. The figure illustrates

that the AC neighborhoods are less isolated and, in fact, are located in the vicinity of the

HD supermarkets in the major commercial districts CD1-CD2. Prices in these AC residential

neighborhoods are likely disciplined by the lower prices in the commercial districts, whereas no

such effect operates in the peripheral neighborhoods.17

Figure 2 plots composite good prices against housing prices, along with a fitted regres-

sion line.18 Prices at AC1, for example, are very high — but are perfectly aligned with that

neighborhood’s affluence level. Prices at the NAP neighborhoods, in contrast, are considerably

higher than what can be systematically associated with their affluence level. While the figure

has only 15 data points, it highlights the message that these peripheral neighborhoods stand

out in terms of the prices charged by their local retailers. In our structural model, we will link

these findings to the presence and effects of spatial frictions.19

Finally, price rankings are quite persistent: the rank correlation of pnt between September

and November 2007 (November 2007 and November 2008) is 0.68 (0.57). This supports our

16AC1, the most affluent neighborhood is, however, usually more expensive than the NAP neighborhoods.17One may wonder why prices at NAP1 are not disciplined by the low prices available at the neighborhood

that lies on its southern border (Pisgat Ze’ev North). Discussions with a resident of that area suggest that thesupermarket at Pisgat Ze’ev North is not particularly attractive for a weekly shopping trip due to its small size.Such issues motivate the inclusion of destination fixed effects in our model of household preferences.

18Commercial districts have a small residential population and therefore we have housing prices there. Thelinear predicted line suggests a positive relationship between composite good and housing prices. But thesmall number of data points (15 observations) and the lack of other controls preclude us from reaching generalconclusions as to whether “the poor pay more.”

19While our framework emphasizes differences in absolute prices charged across neighborhoods with differentincomes, it is also possible to consider income-adjusted prices. For completeness, we also constructed the(G)EKS-Fisher multilateral price index presented in equation (5) in Deaton and Heston (2010). Reassuringly,the correlation between the simple composite price index used in the paper and the GEKS-Fisher index is veryhigh (the correlation coefficient is 0.89, 0.81 and 0.94 in periods 1-3, respectively).

8

focus on spatial, rather than informational frictions.20

2.3 Cross-neighborhood expenditure flow data

We obtained data on consumers’ expenditures from a credit card company. Institutional details

suggest that customers of this company are not different from customers of other companies.

While credit cards are used by 88 percent of the Jewish population, the use of debit cards

is minimal in Israel.21 Our data should therefore be representative of transactions performed

via payment cards. Grocery shopping is, of course, also performed using cash and checks, and

their use may be correlated with important household characteristics. Our econometric model

addresses this measurement problem in detail. We defer discussion of this issue to Section 3.1.

We observe expenditures in supermarkets, grocery stores, bakeries, delicatessen, butcher

stores, wine stores, fruits and vegetables stores and health stores — the type of stores where

our 27 products are likely to be sold — in the same three periods covered by our price data.

The data consist of total expenditures by residents of each origin neighborhood j performed at

each destination neighborhood n where j, n ∈ 1, ..., 46. This results in a 46 by 46 matrix of

expenditure flows between each pair of neighborhoods for each period.

The data were constructed as follows: first, card holders’ neighborhood of residence and

their shopping destination neighborhood were identified via customers’ and stores’ zip codes.22

The expenditure data were then aggregated to the neighborhood level matrix described above.

To be clear, we do not observe data at the individual household or store level.23

Table 5 provides statistics regarding the expenditure data. The most popular commercial

district is CD1 where, on average, 27 percent of expenditures are incurred. CD1 is the top

destination for residents of 16 to 20 of the 46 neighborhoods (depending on the period). CD2

is at a distant second place, although it is quite popular among nearby neighborhoods such as

AC2 (see Appendix Table D7). Most expenditures are not incurred within the home neigh-

borhood, yet home-neighborhood shopping is substantial capturing, on average, 22 percent of

total expenditures. The home destination is the top destination in 12 to 17 cases (depending

20The rise in the average price in the commercial districts in the third period is entirely due to an unexplainedjump in the composite good price at the Romema commercial district (Appendix Table D6). In a robustnesscheck (not reported), we estimated the model without this commercial district, obtaining very similar results.

21Credit cards are also used by 80 percent of the ultra-orthodox Jewish population (https://www.themarker.com/advertising/1.2413558, in Hebrew) On the lack of use of debit cards, see http://www.antitrust.gov.

il/yozma.aspx (in Hebrew).22This required a nontrivial mapping between zipcodes and neighborhoods, where zipcodes can map into

multiple neighborhoods. We employed a “majority rule”: the zip code was mapped to the neighborhood withwhich it has the largest geographical overlap.

23We also observe total expenditures of each origin neighborhood at destinations outside the city. Jerusalemdoes not have a substantial ring of satellite cities providing attractive shopping opportunities. We thereforeconjecture that much of the observed shopping outside the city corresponds to individuals who have a mailingaddress in Jerusalem but do not actually reside in it (e.g., students). We therefore do not use these data inour baseline analysis. Robustness checks (not reported) in which we added the expenditures incurred outsideJerusalem to our model’s “outside option” (see below) yield remarkably close results to the ones reported inSection 3.2.

9

on the period).

As panel B indicates, residents of the non-affluent and peripheral neighborhoods, NAP1-

NAP3, have a higher tendency to shop at their home neighborhood relative to the median

neighborhood whose share of expenditures at home is 0.16. As we have seen in Section 2.2,

this happens despite the fact that shopping “at home” is quite expensive for these residents.

We interpret this as evidence for the importance of spatial frictions. We next explore these

frictions within a structural model of demand for groceries across the city.

3 A structural model of demand in the city

The following subsections present a model of households’ preferences, describe its estimation,

and provide results on price and distance elasticities. We also use the model to calculate the

“Average Price Paid” (hereafter, APP) for each neighborhood of residence. The APP takes into

account the probabilities with which residents of origin j shop in each destination. It therefore

allows for a comparison of the cost of grocery shopping across neighborhoods of residence,

accounting for shopping outside the home neighborhood.

3.1 A model of household preferences

We posit a discrete choice model in which residents in each of the 46 neighborhoods choose

where to perform their grocery shopping. The nested logit framework that we employ is quite

standard. We therefore simply spell out our assumptions and the resulting estimation equation;

a complete derivation of this equation is in Online Appendix F. That appendix also contains

additional discussion and justification for some of our assumptions (in particular, we discuss

our emphasis on spatial rather than information frictions, the single shopping trip assumption,

and additional forms of unobserved taste heterogeneity). Our treatment of measurement error

in the expenditure data, a problem often ignored in applied work, is presented in the text.

Household behavior. A household residing in one of the 46 origin neighborhoods may

shop for the composite good in any one of the n = 1, ..., 15 destinations where composite good

prices are observed in our data. The household may also choose the “outside option” n = 0:

shopping in one of the remaining 31 neighborhoods. The household’s choice maximizes its

utility among these 16 options.

Omitting the time index, the (indirect) utility of household h residing in neighborhood j

from buying the composite good at store s located in neighborhood n is given by

Uhjsn = νc+νj +νn+hpj ·νn+(γ−1 ln yj − ln psn

)·xjα−djn ·xjβ+κ ·hjn+ζhn(σ)+(1−σ)εhjsn

where νc is a constant that shifts the utility from all “inside options” relative to the utility from

10

the outside option (given the standard normalization of the systematic utility of the outside

option to zero). The origin neighborhood fixed effects νj capture utility differences across

origins with respect to this outside option, while the destination fixed effects νn capture quality

differences across destinations. Those may include amenities (parking space, opening hours)

and differences in product variety (i.e., the availability of products other than our basic 27

items). The destination fixed effects are also interacted with the origin neighborhood’s housing

prices, hpj, allowing residents of more affluent neighborhoods to value amenities differently.

The spatial friction — i.e., the utility cost of shopping far away from one’s neighborhood —

is captured via djn, the distance (km) between origin j and destination n, which is interacted

with xj, a vector of origin neighborhood characteristics such as the rate of car ownership. This

friction is further reflected in hjn, a “shopping at home” dummy variable taking the value 1 if

j = n. The parameter κ therefore captures benefits of shopping in the home neighborhood on

top of the implied savings of travel time (and direct travel costs), already captured by β. Put

differently, κ captures a “fixed cost” of shopping outside one’s home neighborhood, possibly

related to the need to drive, or to give up a parking space near home.

Households’ price sensitivity is introduced via the term (γ−1 ln yj − ln psn) · xjα where yj

is the average income in origin neighborhood j and psn is the composite good price at store

s located in destination neighborhood n. This functional form follows Bjornerstedt and Ver-

boven (2016) and implies that, conditional on buying at store s in destination n, the quantity

(units of the composite good) demanded by household h residing in neighborhood j is γyj/psn,

so that expenditure on the composite good is a constant fraction γ of the (representative)

household’s income. The fraction γ drops out of the estimation equation and therefore it

could vary across origin neighborhoods. We note that more sophisticated discrete-continuous

choice models are present in the literature (e.g., Smith 2004, Figurelli 2013). In the context of

our aggregate (neighborhood-level) demand data, we favor this simpler modeling strategy. As

with the distance sensitivity, the price sensitivity is also interacted with origin neighborhood

characteristics.

Finally, the idiosyncratic term ζhn(σ) + (1 − σ)εhjsn, distributed Type-I Extreme Value,

follows the representation of the nested logit model in Berry (1994). The nests are destination

neighborhoods, allowing stores within a neighborhood to be closer substitutes than stores lo-

cated in different neighborhoods. This substitution is governed by the parameter σ that takes

values in the interval [0, 1), with larger values implying stronger within-neighborhood substi-

tutability. The shock εhjsn captures store-level random variation: for example, a household

may particularly value shopping at store s if it is on the way home from work.

We next introduce an assumption, motivated by the nature of our data, regarding symmetry

in the systematic utility provided by stores operating in the same neighborhood:

Assumption 1 Denote by δjsn = νc + νj + νn + hpj · νn − ln psn · xjα− djn · xjβ + κ · hjn the

mean utility level, common to all origin j residents who shop at store s in destination n. Stores

11

within a neighborhood offer identical mean utility levels across households, i.e., δjsn = δjn for

every j, s, n.

Since the only element of δjsn that depends on the store index s is psn, this symmetry

assumption is consistent with a symmetric (within-neighborhood) price equilibrium, i.e., psn =

pn for every store s in neighborhood n. This price symmetry will be consistent with the pricing

model introduced in Section 4.

Assumption 1 implies that stores within a neighborhood are symmetrically differentiated:

they have identical mean utility levels, but offer distinct benefits to individual households

via the idiosyncratic error εhjsn. This assumption therefore allows households who reside in

different streets within the neighborhood to favor the store nearest to them. It also allows for

any other type of horizontal differentiation among stores that is valued idiosyncratically by the

neighborhood’s residents.

The symmetry assumption accommodates the limitation that we observe expenditures at

the neighborhood level rather than at the store level. Tackling such data limitations via a

symmetric differentiation assumption is a familiar strategy in the literature (e.g., Berry and

Waldfogel 1999). This assumption is not very restrictive in our case because stores within a

neighborhood are typically of the same type (e.g., hard discount supermarkets in commercial

districts) and, consequently, as shown in Section 2.2, most of the price variation is across, rather

than within, neighborhoods.24

In Online Appendix F we show that the assumptions spelled out above regarding households’

preferences deliver the following linear equation:

(2) ln

(EjntEj0t

)= νc + νj + (νn + (1− σ) lnLn) + hpj · νn + νt− ln pnt · xjα− djn · xjβ + κ · hjn,

where Ejnt (Ej0t) are total expenditures incurred by residents of origin neighborhood j in

destination n (in the outside option) at time t. Ln is the number of competitors in destination

n which is constant over time.

The left-hand side of (2) contains expenditure shares that are implied by the model but

are measured with error in the data. This error stems from two sources: first, observed prices

pertain to (at most) 27 products, whereas observed expenditures correspond to purchases of

many additional products. Second, we observe credit card expenditures rather than total

expenditures.

Let Ejnt denote expenditures using any payment means on all products sold at the relevant

establishments. Without loss of generality, we can always express expenditures on the 27

24The symmetry assumption would not be needed if we were to use scanner data, since then we would obtainboth price and quantity data at the establishment level. We discussed above, however, the advantages of ourapproach in which we combine establishment-level price data from the statistical authority — that are easilycomparable across space and time — with systematic cross-neighborhood credit card expenditure data. Thelatter provide an efficient coverage of the shopping probabilities characterizing all origin-destination pairs.

12

products using any payment means, denoted by Ejnt, as a proportion of Ejnt, Ejnt = λjntEjnt,

where 0 ≤ λjnt ≤ 1. Similarly, our observed credit-card expenditures on all products, denoted

by Eccjnt, can also be expressed as a proportion of Ejnt, E

ccjnt = τjntEjnt, with 0 ≤ τjnt ≤ 1.

These definitions allow us to map Ejnt, which is derived from the model, into the observed

expenditures Eccjnt via Ecc

jnt = (τjnt/λjnt)Ejnt.

Thus, adding wjnt = ln(τjnt

λjnt

λj0tτj0t

)to the right hand side of (2) allows us to use the observed

ln(Eccjnt/E

ccj0t

)as the dependent variable. The error term wjnt, generated by the mismeasure-

ment of expenditures, presents an identification challenge because the proportionality factors

λ and τ are likely to be correlated with origin and destination neighborhood characteristics.

For example, residents of less affluent origin neighborhoods may have a higher than average

tendency to use cash, and the use of cash may also be more prevalent in certain destinations

(e.g., open fresh produce market). Our strategy for dealing with this endogeneity is to soak up

such tendencies into origin and destination fixed effects via the following assumption:

Assumption 2 Conditional on origin, destination and time fixed effects, wjnt is uncorrelated

with prices and distances.

This assumption implies that the proportionality factors may depend on fixed neighbor-

hood characteristics but can not depend on prices and distance, given these characteristics. It

therefore allows for tendencies of residents of particular neighborhoods to use more or less cash

in certain destinations but assumes that such tendencies are accounted for by the fixed effects.

The panel structure of the data indeed allows us to control for such fixed effects.

Let ujnt be the error from linearly projecting wjnt on a set of origin, destination and time

fixed effects. Assumption 2 allows us to rewrite (2) as

(3) ln(Eccjnt/E

ccj0t) = φc + φj + φn + φt + hpj · vn − ln pnt · xjα− djn · xjβ + κ · hjn + ujnt

where the φ′s are fixed effects. Given the above assumptions, equation (3) is amenable to

consistent estimation via OLS. The observations used consist of all triplets (j, n, t) pertaining to

origin neighborhood j, destination neighborhood n and time period t. The sample size should

therefore equal 46×15×3 = 2, 070 observations, corresponding to expenditure data of residents

of 46 origin neighborhoods at 15 destination neighborhoods over the three time periods. While

the model predicts a positive expenditure share by residents of any origin j at any destination

n, observed expenditures Eccjnt are zero in about 12 percent of all potential observations. We

drop such observations from the sample, reducing its size to 1, 819 observations. The results

are qualitatively robust to substituting a very small number for Eccjnt (online Appendix A).25

25This is not a formally valid correction but one often used in practice (see Gandhi, Lu and Shi 2013 for apartial identification approach).

13

Some of the model’s parameters are not identified. First, the origin, destination and time

dummies (φj, φn, φt) do not identify the utility effects (vj, vn, vt) but rather confound them with

that part of the measurement error wjnt that is correlated with the dummies. An additional

assumption will therefore be required for the computation of elasticities and other quantities of

interest. The consistent estimation of (α, β, κ), however, only requires the assumptions stated

above.

Second, the parameter σ that captures the degree of within-neighborhood competition is

unidentified absent time series variation in the number of competitors in destination n, Ln. This

happens because the fixed effect φn captures the sum of the utility terms vn + (1− σ) lnLn as

well as the linear projection of wjnt on the destination dummy variable. Our practical solution

is to calibrate σ so that it generates reasonable markups given the identified parameters.26

Based on conversations with people familiar with the industry, retail markups of 15-25

percent are reasonable for the type of products studied in this paper.27 Section 4, where the

pricing model is introduced, shows that setting σ = 0.7 yields an average (median) markup

of 22 (20) percent and therefore this is the value chosen for σ. Setting this value to 0.8 or 0.9

instead makes no difference for the qualitative findings in this paper.28 Online Appendix F

provides additional discussion of identification and of our assumptions.

3.2 Estimation results

Table 6 shows OLS estimates of equation (3). We employ 2-way clustering of standard errors at

the origin and destination level, allowing for arbitrary correlation across observations sharing an

origin or a destination. The different specifications control for different sets of fixed effects and

socioeconomic interactions. Across all specifications, the coefficients have the expected signs.

Coefficients on log price and distance (which we entered with a negative sign) are positive,

and so is the coefficient on the “shopping at home” dummy variable, consistent with the high

tendency towards home neighborhood shopping observed in the data (Table 5).

Column (4) includes the full set of origin, destination and period dummies required by

our theory, but without interacting the main regressors with demographics. Both the price

and distance effects have the expected sign and are statistically significant. The inclusion of

destination fixed effects substantially increases the regression’s goodness of fit from 0.38 in

column (2) to 0.66 - 0.78 in columns (3)-(10). This shows the importance of controlling for

26Such an approach has some precedence in the literature. For example, Bjornerstedt and Verboven (2016)calibrate a conduct parameter to generate reasonable markups. We could alternatively pin this parameter downby incorporating supply-side restrictions into the estimation procedure. We favor the calibration as it eliminatesthe need to rely on our pricing model in generating the demand estimates.

27Note that these are markups above marginal cost. They are, therefore, higher than markups over averagecosts, the latter often approximated using information from retailers’ financial reports.

28We also regressed the estimated fixed effects φn on lnLn to estimate 1 − σ. This yields an (impreciselyestimated because of the small number of observations) estimate of σ = 0.81. This estimate is likely to bebiased since vn and the projection of wjnt on vn are likely to be correlated with Ln. Nevertheless, it is somewhatcomforting that the calibrated and estimated values are similar in magnitude.

14

unobserved amenities (e.g., availability of parking, opening hours, product variety etc.).

Columns (5)-(10) then allow for interactions of the price and distance sensitivities with

characteristics of the neighborhood of origin. Households in richer neighborhoods, as proxied

by housing prices, are significantly less sensitive to prices. Distance sensitivity is quite robust

to the inclusion of additional regressors. It is higher in neighborhoods with a large fraction of

elderly residents, though this interaction is not statistically significant. Senior individuals may

face a lower cost of time, but, on the other hand, may find shopping at other neighborhoods

more challenging. The distance sensitivity is smaller in neighborhoods where the share of

residents who own a car, or drive to work, is higher. These effects, however, are only significant

when omitting the “shopping at home” dummy variable in columns (9) and (10).

In column (6), we add the interaction of origin housing prices with destination dummies,

allowing for different valuations of unobserved amenities across socioeconomic classes. The

estimated price coefficient is mildly reduced (from 5.1 to 4.7). Distance coefficients are also

only minimally affected except for the interaction with the percentage of senior citizens.

We adopt column (6) as our baseline specification since it fully and most flexibly controls

for effects that we expect to be important in the household’s decision problem. In particular,

the added interaction term between origin housing prices and destination fixed effects controls

for a broad range of scenarios involving unobserved heterogeneity that could potentially bias

our results (see the discussion in Online appendix F). Overall, estimates in columns (7) – (10)

are very close to the baseline specification in column (6).29

Elasticities. The economic implications of these estimates are captured in price and dis-

tance elasticities. The own-price elasticity faced by store s in neighborhood n is (Appendix

F):

(4) ηsn,p =psnQsn

∂Qsn

∂psn= −

J∑j=1

Qjn

Qn

[1 + xjα

(1

1− σ− σ

(1− σ)Ln− πjnLn

)]

where Qsn is the total demand at store s located in neighborhood n. πjn is the probability

that a resident from origin j shops in neighborhood n. The total demand faced by all retailers

in neighborhood n is denoted by Qn, whereas Qjn is the part of this demand generated by

residents of origin j. The demand elasticity faced by the store is therefore a quantity-weighted

average of origin-specific elasticities, where the weights depend on the fraction of the retailers’

demand generated by residents of those origins.

The semi-elasticity of Qjn with respect to the distance between j and n is

29In column (7), the price coefficient changes because its interaction with housing prices is omitted. However,the own price elasticities implied by columns (6) and (7) are nearly identical, differing by 5% on average.Moreover, interacting price with family size (not reported) yields an insignificant effect and does not alter theother coefficients.

15

ηjn,d =1

Qjn

∂Qjn

∂djn= −xjβ (1− πjn) ,

measuring the percentage change in demand from residents of neighborhood j at destination

n 6= j in response to a 1 km increase in the distance between these neighborhoods.

Estimating the elasticities requires the estimated parameters obtained above (including the

calibrated σ), an estimate of the choice probabilities πjn, and data on the number of stores in

destination n, Ln. We set Ln equal to the number of the neighborhood’s supermarkets in 2008.30

We do not directly count grocery stores, or other non-supermarket retail establishments, as these

are not close substitutes to supermarkets for the purpose of the weekly grocery shopping (e.g.,

because of limited availability of items). Nonetheless, to partially take these additional store

formats into account, we add the value of one to Ln in residential neighborhoods, while keeping

it equal to the number of supermarkets in the commercial districts. This modification results

in more reasonable estimated margins and has a negligible effect on the qualitative findings of

the counterfactual analyses reported in Section 5.

Estimation of the choice probabilities πjn is complicated by the fact that these depend on

the mean utility levels and, therefore, on the utility fixed effects (ν) which are not identified.

Appendix F shows that the mean utility levels are identified under the following assumption:

Assumption 3 For each origin j, the ratioτjnλjn

is identical for all destinations n.

This assumption implies that choice probabilities are equal to the observed expenditure

shares. This is clearly weaker than simply assuming their equality which amounts to ignoring

the measurement error altogether. We stress that Assumption 3 is not required for the consistent

estimation of the parameters α, β, κ. Our framework therefore clarifies the different sets of

assumptions that can be used to accomplish different goals in the presence of measurement

error in the expenditure data.

Employing the leading specification (column 6 of Table 6) we estimate price elasticities

for each destination, and distance semi-elasticities for each origin-destination pair. Table 7

presents estimates for the last period, November 2008, which are nearly identical to the average

over the three periods. The average (median) store-level own price elasticity ηsn,p is 4.82

(4.95) in absolute value. The individual estimates are tightly distributed around the mean.

Recalling that close substitutes are often available in the form of other stores within the same

neighborhood, this relatively-elastic demand seems reasonable. Increasing σ to 0.8 generates a

higher mean price elasticity of 6.43 but, as reported in online Appendix B, it makes no difference

30The number of supermarkets is shown in the last column of Appendix Table C2. This value includes allsupermarkets, not just those where prices were sampled. A specific issue arises with respect to the open marketof Mahane Yehuda where many small sellers – open stalls – are present. To retain internal consistency, we setLn = 2 in that location (because there is a small supermarket in the neighborhood). Using different valuesaffects the margins for retailers in this specific neighborhood, but does not affect the qualitative findings.

16

in terms of our qualitative conclusions.31 Online Appendix A presents the robustness of the

estimated elasticities to alternative computations of the price of the composite good discussed

in Section 2.2.

The average (and median) distance semi-elasticity ηjn,d is 0.35 in absolute value implying

that a 1 km increase in the distance between an origin j and a destination n decreases demand

by residents from j at n by 35 percent. Spatial frictions are, therefore, a first-order consid-

eration affecting households’ choices, consistent with the anecdotal evidence surveyed in the

Introduction.

To assess the price-distance trade-off, we consider residents of location j who shop at des-

tination n. The maximum percentage price increase these consumers are willing to accept for

destination n to become one kilometer closer to them is 100 (exp (xjβ/xjα))− 1). The median

of this estimated quantity over the 46 origin neighborhoods is 24.5 percent indicative, again, of

a substantial spatial dimension in households’ preferences.

3.3 Retail price differences revisited: the Average Price Paid

In Section 2.2 we showed that prices in non-affluent, peripheral neighborhoods (NAP) were often

higher than prices in more affluent but centrally located neighborhoods. This, however, does

not necessarily imply that households in NAP neighborhoods pay more for groceries because

they may not shop in their neighborhood of residence. Our estimated model allows us to revisit

this issue.

Recall that, given Assumption 3, the probability that a resident from neighborhood j buys

the composite good in neighborhood n, πjn, can be computed directly from the expenditure

data. We can therefore compute the Average Price Paid (APP) for residents of neighborhood

j: pAj ≡∑N

m=0 πjmpm. This weighted average takes into account the probabilities with which

residents of origin j shop in each destination n. It therefore allows a comparison of the cost

incurred by residents of different origin neighborhoods, taking into account their shopping

behavior.32

Figure 3 plots the Average Price Paid against housing prices in each of the 46 neighborhoods

in November 2008 (along with a linear predicted line). Interestingly, the centrally-located, af-

fluent neighborhood AC1 has an APP which is perfectly explained by its affluence level. This

was also true for the price charged by retailers operating in this neighborhood, as we saw in

Figure 2, which we refer to as the “posted” price. The three non-affluent, peripheral neigh-

borhoods NAP1-NAP3 have an APP that lies above the predicted line, although this gap is

less pronounced relative to the gap observed in Figure 2 for the posted prices charged in these

31Further, increasing σ to 0.9 makes demand even more elastic, which is intuitive, but once again does notaffect our qualitative conclusions. Details are available from the authors upon request.

32This requires also an estimate of the price of the composite good at the outside option neighborhoods, p0,which is unobserved. These outside option neighborhoods are residential neighborhoods where we believe mostshopping opportunities are at expensive grocery stores. We therefore set p0 as the price charged in NAP3: theperipheral, non-affluent neighborhood that launched the consumer boycott in 2014.

17

neighborhoods. This suggests that despite the ability to shop outside the home neighborhood,

residents of these NAP neighborhoods still pay more, on average, than what could be explained

by their affluence level, consistent with the spatial friction. In 8 out of the 11 residential neigh-

borhoods with valid prices, the APP is substantially lower than the observed price, reflecting

the savings afforded to households by shopping outside their home neighborhood (when pAj is

higher than pj, the difference is small).

It is of interest to compare pAj with the minimum price across all 15 neighborhood, which

would have been the price actually paid if households were to determine their shopping desti-

nation based on price only (ignoring equilibrium effects). The APP is, on average, 12.2 percent

higher than this minimum price (the range being between 3.7 and 21.2 percent). This number

reflects the monetary value of the spatial frictions faced by households (captured in the model

via β and κ) as well as their preferences for specific shopping destinations (captured by vn and

the idiosyncratic terms). It also provides a rough indicator of the maximal extent to which

prices can be expected to decline were these frictions removed.

Importantly, the Average Prices Paid at the peripheral, non-affluent neighborhoods NAP1-

NAP3 are higher than those faced by residents of more affluent neighborhoods that are located

closer to the commercial areas, AC2-AC3. Moreover, a strong, positive relationship is depicted

in Figure 4 between the APP pAj and distance to the main commercial district CD1. This is

yet another manifestation of the role played by spatial frictions in determining the variation in

the cost of grocery shopping across households.

We next introduce a model of pricing decisions and employ it, along with the estimated

demand system, to tease out the effect of spatial frictions on the cost of groceries for residents

of neighborhoods across the city, and, specifically, in the NAP neighborhoods. A supply model

is needed in order to allow retailers to adjust their pricing decisions in counterfactual scenarios.

This will be necessary, for example, when calculating the response of equilibrium prices across

the city to a city-wide reduction in the travel cost.

4 Retail supply: pricing decisions

Our model for the supply side of the market is summarized in the following assumption:

Assumption 4 (i) Each neighborhood n features Ln retailers that share a constant-in-output,

symmetric marginal cost cn. (ii) Retailers across the city engage in Bertrand competition: each

store s in each neighborhood n chooses its price psn simultaneously. (iii) A unique, interior

Nash equilibrium in prices exists. (iv) Equilibrium prices can differ across neighborhoods, but

satisfy within-neighborhood symmetry: psn = pn at each store s in each neighborhood n.

Part (i) of Assumption 4, along with the symmetric differentiation assumption from the

demand model (Assumption 1) naturally allow us to focus on the (within-neighborhood) sym-

metric price equilibria assumed in part (iv).

18

The standard first order conditions from this model (see online appendix F) imply that

the margin (pn − cn)/cn is inversely related to the own-demand elasticity in (4). Margins are

therefore intuitively tied down to the model’s primitives. Specifically, the margin garnered by

neighborhood-n retailers increases in πjn because it reflects neighborhood-j residents’ tendency

for shopping at n. This effect is mediated via demographics: it is stronger, the lower is the

sensitivity of residents of j to price, reflected in a high value of xjα. The effect also increases

in the share of sales by neighborhood n retailers to households from neighborhood j, Qjn/Qn.

In a residential neighborhood n, the term Qnn/Qn – the fraction of the sales by retailers located

at n made to residents of the same neighborhood – is usually large and will be dominant in

determining the margin at n. If n is a peripheral neighborhood, πnn will be large (Table 5) —

implying an inelastic demand working in the direction of increasing the retail margin.

Table 8 displays the estimated costs and margins by neighborhood in November 2008, the

time period in which we conduct the counterfactual analyses. Very similar quantitative and

qualitative patterns obtain when averaging over the three time periods. Using the baseline

value σ = 0.7, the average (median) estimated margins are 22 (20) percent. Conversations with

people familiar with the retail industry in Israel suggest that this is a reasonable margin given

the type of products considered in this paper. Indeed, this value for σ was chosen precisely for

this reason (see Section 3.1). We also compute margins assuming σ = 0.8, generating somewhat

lower margins but the same qualitative counterfactual conclusions (online Appendix B).

Margins in residential neighborhoods are generally higher than those in the large commercial

districts. Our model attributes this to both spatial frictions and to low within-neighborhood

competition in residential areas. Furthermore, marginal costs at the non-affluent, peripheral

neighborhoods NAP1-NAP3 are particularly high. This may reflect the cost of transporting

goods into more remote locations, and the lower operational economies of scale obtained at the

relatively smaller supermarkets located there.33

5 Counterfactuals: the impact of spatial frictions

We perform counterfactual analyses to assess the role of spatial frictions in generating the city’s

price equilibrium. In a first exercise we make travel less costly, and in a second one we increase

the appeal of the commercial districts. Intuitively, both changes should make households more

willing to shop in the commercial districts. In a third exercise, we artificially increase the

number of retailers in residential neighborhoods. We examine the impact of these changes on

(i) the equilibrium posted prices, i.e., the prices charged by retailers in each neighborhood (ii)

the probabilities with which residents of each origin neighborhood shop at each destination, and

(iii) the Average Price Paid (APP) by residents of each neighborhood taking these probabilities

33We do not consider multi-store pricing by chains that operate supermarkets in both residential and commer-cial districts. Because these chains’ pricing in the commercial districts is strongly constrained by the presenceof hard discounters that operate only there, we expect this issue to have little impact on our findings.

19

into account.34

We use the baseline estimates from column 6 in Table 6 and σ = 0.7. Online Appendix B

reports counterfactual results using the value σ = 0.8, delivering very similar results (as is also

the case for σ = 0.9, with details available from the authors upon request). Online Appendix

G provides technical details on the computation of the counterfactual equilibria.

Table 9 summarizes the impact on the posted prices charged by retailers across the city.35

The first column corresponds to the observed posted prices whereas the other columns report

the predicted percentage changes to those prices in each counterfactual experiment.

The first experiment reduces the distance disutility by half: that is, we add 0.5djnxjβ to the

utility garnered by residents of each origin j from shopping in each destination n. The second

experiment reduces spatial frictions even further by reducing, in addition, the preference for

shopping at home parameter κ by 50 percent. As Table 9 shows, these two experiments reduce

the median (across the 15 neighborhoods with observed prices) prices by 0.7 to 1 percent.

Median posted prices within the 11 residential neighborhoods with observed prices are reduced

by only one half of a percent. These are quite modest declines. Examining our non affluent,

peripheral neighborhoods, NAP1-NAP3, we see that prices are reduced by 0.5 percent in NAP3

(where the boycott took place), but actually increase in NAP1 and NAP2.

At first glance these increases appear puzzling since a reduction in spatial frictions should

enhance competition across the city, exerting a substantial downward pressure on retail prices

in the residential neighborhoods, and certainly in the peripheral ones. Why, then, do prices

decline only mildly or even increase? The answer lies in the changes in the composition of

demand faced by retailers in these neighborhoods. As travel becomes less costly, households

that continue shopping in the expensive residential neighborhoods are those with very large

idiosyncratic shocks favoring shopping there. Retailers in these neighborhoods therefore face

a less elastic demand prompting them to raise prices, providing a countervailing force that

diminishes, and sometimes even offsets, the competitive force.36

A similar picture arises when improving the amenities in the city’s main commercial district,

CD1, and at the two major ones, CD1-CD2, respectively. This is performed by increasing the

destination fixed effects νn associated with each such district by one standard deviation.37 This

may correspond to various improvements in the shopping experience in these districts: for

example, the city may improve the physical infrastructure by setting up large parking spaces

34The model allows us to compute the impact on welfare but as our exercises directly affect utility parameterswe find this less appealing.

35Online Appendix E provides complete neighborhood-specific results for all the counterfactuals.36This is similar to the observed “generic drug paradox” that occurs when many, but not all, consumers

switch to newly available generic drugs but prices among the incumbent (brand) drugs do not decline and evenrise (Griliches and Cockburn, 1994).

37Given that νn is unidentified due to measurement error, we use the standard deviation of φn, the fixed effectthat confounds νn with the measurement error effect, instead. One standard deviation of the distribution of φnmay be greater than one standard deviation of the distribution of νn. This issue, however, does not drive ourfindings. We obtain very similar qualitative findings by adding one half of a standard deviation of φn instead.

20

at the entry points to the commercial district with a convenient shuttle service.38 Boosting the

utility of shopping at n can make the citywide grocery market more competitive. But the same

countervailing force applies here so, again, only mild price reductions are observed (median

price declines of 0.1 – 0.8 percent in residential neighborhoods, and of 0.1 – 0.9 percent in

NAP1-NAP3).

Finally, the last column considers the effect of exogenously increasing the number of competi-

tors in each residential neighborhood n by 1. The median posted price decline across residential

neighborhoods is 3.4 percent, whereas the price declines in the NAP1-NAP3 neighborhoods are

1.3 – 3.5 percent. Such additional entry, however, may be associated with substantial social

opportunity costs due to zoning restrictions and lack of space. A price reduction of about 3.5

percent may not be large enough to justify such costs.

Whereas equilibrium posted prices respond very mildly, households’ shopping behavior

changes markedly: many households switch to shopping in the affordable commercial districts.

This is captured by the APP, the weighted average of prices paid by residents of each neighbor-

hood, taking into account their probability of shopping in the various destinations. Table 10

summarizes the counterfactual changes to the APP faced by residents of the eleven residential

neighborhoods where prices are observed.39

The first column of Table 10 reports the APP faced by residential neighborhoods in the

observed equilibrium. The percentage reduction in APP faced by neighborhood residents fol-

lowing a reduction in the spatial friction is much larger than the corresponding percentage

reduction in prices charged in the neighborhood as reported in Table 9. As the top row of

Table 10 indicates, in the first four experiments the median (over residential neighborhoods)

APP falls by 1.8 - 5.6 percent, versus the 0.1 - 0.8 percent fall in median posted prices shown

in Table 9.

The Average Price Paid, therefore, responds much more strongly than the posted price to

changes in spatial frictions. A median average reduction of 5.6 percent is quite substantial

because, as remarked in Section 3.3, the average difference between the APP and the minimum

price in the observed equilibrium is about 12.2 percent which can be taken as an upper bound

to the price effect in our counterfactual exercises.

In short, while Table 9 considers only the impact on equilibrium prices charged in different

locations, the Average Prices Paid in Table 10 take also into account the changes in shopping

patterns induced by the changes in parameters. This is evident in Figure 5 that compares the

probability of shopping at CD1, the city’s most important commercial district, in the observed

equilibrium, to the same probability under the counterfactual that improves amenities in that

38Interestingly, the city of Jerusalem recently announced plans to improve the main commercial district, CD1,exactly along these lines: “The plan: the Talpiot industrial zone expected to undergo a revolution over the nextdecade,” Kol Hair (April 15, 2016).

39For completeness, Appendix Table E2 shows the impact on the APP for all 46 neighborhoods, deliveringthe same qualitative conclusions. We favor presenting here results for the 11 residential neighborhoods whereprices are observed to facilitate comparison with the impact on posted prices displayed in Table 9.

21

district. The probability of shopping at CD1 increases for residents of all neighborhoods, and

substantially more for those located in the periphery. While the price charged at CD1 increases

slightly, it is still low, and, as a consequence, the Average Prices Paid decline considerably.

Viewed through the lens of its effect on the APP, the benefits from decreasing spatial

frictions to the average resident of the three disadvantaged neighborhoods NAP1-NAP3 are

substantial. When amenities at CD1 are improved, the APP incurred by residents of NAP3 —

the neighborhood where the first boycott took place — drops by a substantial 7 percent (Table

10), whereas the posted price charged by the retailers at NAP3 dropped by 0.6 percent only

(Table 9). Average Prices Paid at NAP1 and NAP2 drop by 2.2 and 6.6 percent, respectively,

whereas posted prices charged by the retailers in both of these neighborhoods only drop by at

most 0.1 percent.

Discussion. Evaluating the impact of a reduction in spatial frictions by considering only

the effect on posted prices would be misleading: it would suggest very mild benefits, if at all.

In contrast, the analysis that also considers the impact on shopping probabilities, embedded

into the computation of the Average Price Paid, suggests substantial reductions in the cost of

grocery shopping. This point applies to residential neighborhoods in general, and not only to

the disadvantaged ones.

Reducing spatial barriers or improving amenities at commercial districts may confer substan-

tial benefits to the average resident of the peripheral, non-affluent neighborhoods.40 However,

because prices charged at the NAP neighborhoods barely change, or even increase, a reduc-

tion in spatial frictions does little to benefit residents with limited mobility (e.g., the young

families described as having no time to shop, or the elderly). Those residents will continue

to pay the expensive prices in their home neighborhood. Importantly, these are precisely the

household segments identified by the boycott organizers as experiencing the most harm from

the price differentials in the first place. Alleviating the cost of living for such individuals may

therefore require more targeted relief programs.41 Increased competition, in the form of gradual

improvements to the city’s infrastructure, will fail to provide relief to such populations.

6 Summary and conclusions

This paper uses a unique dataset on prices in spatially-differentiated neighborhoods within a

large metropolitan area, and on the distribution of expenditures across these neighborhoods,

to explore the determinants of price differentials and shopping patterns within the city. We

document that retailers at several peripheral, non-affluent neighborhoods often charge higher

prices than retailers located in more centrally located, affluent neighborhoods.

40A caveat to this statement is that the attractiveness of a location vn is probably endogenous and mightchange if it experiences a substantial increase in shopping activity. Addressing this would require a model ofretailers’ choice of amenities which is beyond the scope of this paper.

41One possibility would be to add location and mobility, in addition to income, to the criteria determiningeligibility into welfare programs, e.g., food stamps.

22

Using an estimated structural model of demand and supply, we establish that spatial fric-

tions play an important role in generating these patterns. Our counterfactual analysis reveals

that alleviating spatial frictions brings substantial benefits to the average resident of periph-

eral neighborhoods. This operates by improving access to hard discount supermarkets located

in the commercial districts. The prices charged in the residential neighborhoods themselves,

however, do not decline much, and sometimes even increase further exacerbating the cost of

living conditions for households with limited mobility.

Our simple model can be extended in future work to accommodate multi-store pricing by

retail chains, or more complicated demand systems. The parsimony of the model presented

here has the important benefit that the demand model can be estimated via linear regressions.

The model is capable of producing reasonable predictions that are consistent with institutional

details and anecdotal evidence regarding the nature of retail spatial competition within an

urban setting. We view the paper as a step toward a better understanding of the role played

by spatial frictions in determining the cost of grocery shopping across a city.

Our analysis reveals the benefits from enhanced household mobility. It also reveals that

better mobility, while strengthening competition across retailers, will not necessarily lead to

lower prices in periphery neighborhoods, perhaps motivating more targeted policies aimed at

alleviating the cost of living for low-mobility households residing there.

References

[1] Adams, B. and K.R. Williams (2019): “Zone Pricing in Retail Oligopoly,” American Eco-

nomic Journal: Microeconomics, 11(1), pages 124-156.

[2] Aguiar, Mark and Erik Hurst (2007) “Life-Cycle Prices and Production,” American Eco-

nomic Review, American Economic Association, vol. 97(5), pages 1533-1559

[3] Aguirregabiria, V. and J. Suzuki (2015): “Empirical Games of Market Entry and Spatial

Competition in Retail Industries,” CEPR Discussion Paper No. 10410

[4] Baye, M.R., J. Morgan, and P. Scholten (2006): “Information, Search, and Price Dis-

persion,” in: Terrence Hendershot (Editor), Handbook on Economics and Information

Systems, Elsevier.

[5] Berry, S.T. (1994): “Estimating Discrete-Choice Models of Product Differentiation,”

RAND Journal of Economics, 25(2): 242-262

[6] Berry, S., Levinsohn, J., and A. Pakes (1995): “Automobile Prices in Market Equilibrium,”

Econometrica, 63(4): 841-90

[7] Bjornerstedt, J., and F. Verboven (forthcoming in 2016): “Does Merger Simulation Work?

Evidence from the Swedish Analgesics Market,” American Economic Journal: Applied

Economics.

23

[8] Caplin, A., and B. Nalebuff (1991): “Aggregation and Imperfect Competition: On the

Existence of Equilibrium,” Econometrica, 59(1): 25-59

[9] Caplovitz, D. (1963). The poor pay more. New York: Free Press of Glencoe.

[10] Chintagunta, P., J.-P. Dube, and V. Singh (2003): “Balancing Profitability and Customer

Welfare in a Supermarket Chain,” Quantitative Marketing and Economics, 1(1): 111–147.

[11] Chung, C., and S.L. Myers (1999). “Do the poor pay more for food? An analysis of grocery

store availability and food price disparities.” Journal of consumer affairs, 33(2), 276-296.

[12] Davis, P. (2006): “Spatial competition in retail markets: movie theaters,” The RAND

Journal of Economics, 37(4), 964–982

[13] Davis, R.D., Dingel, J.I., Monras, J., and E. Morales (2019), “How Segregated is Urban

Consumption?,” Journal of Political Economy, 127(4), 1684-1738

[14] Dubois, P., and S. Jodar-Rosell (2010): “Price and Brand Competition between Differen-

tiated Retailers: A Structural Econometric Model,” Mimeo

[15] Ellickson, P.B., Grieco, P.L. and O. Khvastunov (2016), “Measuring Competition in Spatial

Retail”, Mimeo.

[16] Figurelli, L. (2013): “Store Choice in Spatially Differentiated Markets,” Mimeo

[17] Gandhi, A., Z. Lu, and X. Shi (2013): “Estimating Demand for Differentiated Products

with Error in Market Shares,” Mimeo

[18] Griliches, Z., and I. Cockburn (1994), “Generics and New Goods in Pharmaceutical Price

Indexes,” American Economic Review, 84(5), 1213-32.

[19] Handbury, J. (2013) “Are Poor Cities Cheap for Everyone? Non-Homotheticity and the

Cost of Living Across U.S. Cities”, Mimeo.

[20] Hotelling, H. (1929): “Stability in competition,” Economic Journal 39, 41-57

[21] Houde, J.-F. (2012): “Spatial differentiation and vertical contracts in retail markets for

gasoline,” American Economic Review, 102(5): 2147-82

[22] Kurtzon, G., and R. McClelland (2010): “Do the Poor Pay More Store-by-Store?,” BLS

Working Paper 437

[23] MacDonald, J. M., and P. E. Nelson (1991). “Do the poor still pay more? Food price

variations in large metropolitan areas.” Journal of Urban Economics, 30(3), 344-359.

[24] McManus, B. (2007): “Nonlinear pricing in an oligopoly market: The case of specialty

coffee,” RAND Journal of Economics 38, 512-532

[25] Miller, N.H. and M. Osborne (2014):“Spatial differentiation and price discrimination in the

cement industry: evidence from a structural model.” The RAND Journal of Economics

45(2): 221-247

[26] Moretti, E. (2013). “Real Wage Inequality,” American Economic Journal: Applied Eco-

nomics, 5(1): 65-103.

[27] Nocke, V., and N. Schutz (2016), “Multiproduct-Firm Oligopoly: An Aggregative games

Approach,” CEPR DP 11539m a structural model,” RAND Journal of Economics 45(2):

221-247

24

[28] Pinske, J., M. E. Slade, and C. Brett (2002):“Spatial price competition: A semiparametric

approach,” Econometrica 70(3), 1111-1153

[29] Richburg-Hayes, L. (2000). “Do the poor pay more? An empirical investigation of price

dispersion in food retailing”. Princeton Dept. of Econ., Industrial Relations Working Paper,

(446).

[30] Salop, S. (1979): “Monopolistic competition with outside goods,” Bell Journal of Eco-

nomics 10, 141-156

[31] Smith, H. (2004): “Supermarket Choice and Supermarket Competition in Market Equi-

librium,” Review of Economic Studies, 71(1), 235–263

[32] Thomadsen, R. (2005): “The effect of ownership structure on prices in geographically

differentiated industries,” RAND Journal of Economics 36, 908-929

25

7 Tables and Figures

Table 1: Distribution of demographics across Jerusalem neighborhoods

Variable N mean sd min p25 p50 p75 max

Population (000s) 46 15.0 5.3 6.2 10.5 13.9 18.3 28.7Households (000s) 46 4.4 1.6 2.1 3.3 4.2 5.3 8.8Average household size 46 3.4 0.9 1.9 2.8 3.3 4.1 6.1Housing prices (000s) 46 13.4 3.0 8.8 11.5 13.3 15.2 21.1% Driving to work 46 39.7 18.6 7.5 23.8 47.2 55.3 68.1% Car ownership 46 48.9 22.9 6.9 34.4 59.2 65.9 89.3% Senior citizens 46 10.6 4.9 1.1 7.5 10.2 14.4 25.6

Notes: Housing prices = the 2007-2008 average price per square meter in thou-sands of NIS, Driving to work = percentage of individuals above 15 years ofage who used a private car or a commercial vehicle (as a driver) as their mainmeans of getting to work in the determinant week. Car ownership = percentageof households using at least one car. Senior citizens = percentage of individualsabove age 65. Source: Central Bureau of Statistics (CBS).

Table 2: Distribution of distance between Jerusalem neighborhoods (in km)

Variable N mean sd min median max

Distance to City center 46 4.3 2.3 0.6 4.3 9.2Distance to Commercial District 1 (CD1) 46 5.7 2.9 0.0 5.4 13.2Distance to Commercial District 2 (CD2) 46 6.0 2.5 0.0 5.8 12.0Mean distance to all other neighborhoods 46 6.1 1.6 4.2 5.8 10.8

Notes: Statistics of the distribution of distances in kilometers between each neighbor-hood and 1) the city center, 2) the two prominent commercial centers CD1 and CD2,and 3) all other neighborhoods. Source: CBS

26

Table 3: Number of sampled stores and of observed products

# sampled stores # observed products # supermarkets

Sep2007 Nov2007 Nov2008 Sep2007 Nov2007 Nov2008

A. Statistics over all 26 neighborhoods where prices were collected

Mean 2 2 2 18 17 17 1Min 0 0 0 0 0 0 0Max 10 10 9 27 27 27 5

Total 54 55 51 29

B. Particular neighborhoods of interest

NAP1 1 1 1 27 27 27 1NAP2 2 2 2 27 27 27 2NAP3 3 2 2 27 26 26 1

AC1 2 2 2 24 25 24 1AC2 3 3 3 27 27 27 2AC3 1 1 1 26 25 23 1

CD1 7 7 7 27 27 27 5CD2 3 3 3 27 27 26 3

Notes: Statistics regarding the number of stores and products sampled across the city. CD, NAPand AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neigh-borhoods (see text). The number of supermarkets (most-right column) includes all supermarketsin the neighborhood, not just those where prices were sampled. Some neighborhoods are not sam-pled in all three periods and therefore have zero stores in at least one period. Many of the tenstores sampled in the main open market are fresh produce stalls.

27

Table 4: Price of composite good across Jerusalem neighborhoods

A. Statistics over 15 neighborhoods with at least 21 observed price items

Prices (NIS)

Sep-07 Nov-07 Nov-08

Mean residential (11) 7.40 7.19 7.92Mean commercial (4) 6.91 6.81 7.46

Min residential (11) 6.23 6.56 7.36Min commercial (4) 6.33 6.15 6.89

Max residential (11) 8.01 7.61 8.52Max commercial (4) 7.45 7.30 8.69

B. Particular neighborhoods of interest

Prices (NIS) Price rank (lowest=1, highest=15)

Sep-07 Nov-07 Nov-08 Sep-07 Nov-07 Nov-08 Mean rank

NAP1 7.15 7.31 8.01 6 10 10 9NAP2 7.54 7.39 8.14 10 14 11 12NAP3 7.80 7.36 8.19 14 12 13 13

AC1 8.01 7.27 8.52 15 8 14 12AC2 7.55 7.61 7.85 11 15 8 11AC3 7.68 7.06 7.76 13 7 7 9

CD1 6.33 6.15 6.89 2 1 1 1CD2 7.45 7.30 7.07 9 9 2 7

C. Gains from travel: distribution of savings (%) from shopping at the cheapest location

Percentile 10% 25% 50% 75% 90%

Savings 2.58 9.66 14.01 16.39 19.13

Notes: Panel A displays composite good price statistics over the 15 neighborhoods where the price could be com-puted using at least 21 observed products (see text). Panel B presents values for particular neighborhoods whereCD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neighbor-hoods (see text). The last column in panel B presents the neighborhood’s mean price rank over the three sampleperiods. For example, the CD1 commercial district has a mean rank of 1, i.e., it is, on average, the cheapest lo-cation. Panel C shows the distribution of savings, in percentage terms, from shopping at the cheapest locationrather than at the home neighborhood using data for all 15 neighborhoods where the price could be computedusing at least 21 observed products over the sample period. For example, the median savings are 14.01%.

28

Table 5: Credit card expenditure flows

A. Statistics over all 46 neighborhoods

Fraction spent atOwn neighborhood CD1 CD2

Mean 0.22 0.27 0.06Median 0.16 0.19 0.03Min 0.00 0.01 0.01Max 0.76 0.76 0.41

B. Neighborhoods of interest

Fraction spent atOwn neighborhood CD1 CD2

NAP1 0.25 0.03 0.02NAP2 0.42 0.18 0.04NAP3 0.33 0.31 0.05

AC1 0.44 0.19 0.03AC2 0.14 0.16 0.18AC3 0.00 0.65 0.02

Notes: The table provides statistics (averaged over the sample period) on thefraction of expenditures spent at the own neighborhood and at the CD1 and CD2commercial districts. NAP and AC correspond to Non-Affluent Peripheral, andAffluent Central neighborhoods (see text).

29

Table 6: Estimates of utility function parameters

Variable (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

ln (price) 8.768 9.283 1.691 1.725 5.065 4.727 1.630 4.730 5.865 4.646(5.788) (5.491) (.763) (.749) (1.421) (1.304) (.774) (1.302) (1.537) (1.333)

ln(p) × housing prices -0.253 -0.232 -0.232 -0.315 -0.228(.083) (.078) (.078) (.091) (.079)

Distance 0.272 0.365 0.197 0.334 0.393 0.423 0.423 0.411 0.471 0.487(.049) (.072) (.036) (.045) (.13) (.12) (.12) (.119) (.116) (.107)

Distance × seniors 0.002 0.004 0.004 0.004 0.004 0.006(.004) (.007) (.007) (.006) (.005) (.007)

Distance × driving to work -0.002 -0.003 -0.003 -0.003(.002) (.002) (.002) (.001)

Distance × car ownership -0.002 -0.003(.001) (.001)

Shopping at home 2.489 1.723 3.035 2.089 1.977 1.890 1.889 1.910(.426) (.526) (.397) (.41) (.435) (.426) (.426) (.424)

Fixed origin effects NO YES NO YES YES YES YES YES YES YESFixed destination effects NO NO YES YES YES YES YES YES YES YESFixed period effects YES YES YES YES YES YES YES YES YES YESDestination× housing p. NO NO NO NO NO YES YES YES YES YES# observations 1819 1819 1819 1819 1819 1819 1819 1819 1819 1819R2 0.243 0.382 0.657 0.775 0.776 0.784 0.783 0.783 0.762 0.770

Notes: The table reports estimates of equation (3). The price and distance variables were entered with a negative sign inthe regression so that the estimates in the table are estimates of α and β. The second row reports the coefficient for theinteraction of the natural log of prices at the destination with housing prices at the origin. The distance variable is inter-acted with the origin neighborhood’s share of senior citizens, its share of residents who drive to work, and its car ownershipshare. The term “Destination× housing p.” refers to the interaction of destination neighborhood fixed effects with housingprices at the origin neighborhood. Standard errors in parentheses are 2-way clustered at the origin and destination levels.

30

Table 7: Distribution of estimated elasticities across neighborhoods (absolute value)

Own price elasticity

σ mean sd min p10 p25 p50 p75 p90 max N

0.7 4.82 0.92 3.00 3.86 3.99 4.95 5.87 5.95 6.13 150.8 6.43 1.37 3.78 5.01 5.31 6.54 7.94 8.32 8.47 15

Distance semi-elasticity

mean sd min p10 p25 p50 p75 p90 max N0.35 0.06 0.06 0.28 0.31 0.35 0.39 0.42 0.45 690

Notes: All elasticities computed given the baseline demand estimates (col-umn 6 of Table 6) for November 2008. Own price elasticities are presentedfor alternative values of σ, while distance semi-elasticities are at the neigh-borhood level and do not depend on σ. The table shows statistics for elas-ticities computed at each of the 15 destinations where the composite goodprice could be computed using at least 21 items. Price elasticities werecomputed for each of these 15 destinations. Distance semi-elasticities werecomputed for each of the 690 (46x15) origin-destination pairs.

Table 8: Estimated costs and margins

σ=0.7 σ=0.8

p c (p-c)/p c (p-c)/p

Average (all) 7.80 6.11 0.22 6.52 0.16

Median (all) 7.85 6.07 0.20 6.44 0.15

Median CD1-CD2 6.98 5.69 0.18 6.04 0.13

Median residential 7.87 6.10 0.21 6.44 0.16

Median NAP1-NAP3 8.14 6.66 0.17 7.00 0.13

Median AC1-AC3 7.85 5.88 0.25 6.37 0.19

Notes: The table reports the composite good price (p), marginalcost (c), and price-cost margin (p-c)/p in each destination neigh-borhood where the composite good price could be computed us-ing at least 21 items in November 2008. Costs and margins arereported under two alternative values for the correlation param-eter σ. CD, NAP and AC correspond to Commercial Districts,Non-Affluent Peripheral, and Affluent Central neighborhoods (seetext).

31

Table 9: Counterfactual changes to posted prices

Retail location Observed posted price Reduced travel disutility Improved amenities Additional entry

(NIS) Distance Distance & κ CD1 CD1-CD2

Average (all) 7.80 -0.9% -1.4% -0.3% -0.3% -2.6%

Median (all) 7.85 -0.7% -1.0% 0.0% -0.3% -3.0%

Median CD1-CD2 6.98 -0.7% -0.6% 0.4% 0.3% 0.0%

Median residential 7.87 -0.5% -0.5% -0.1% -0.8% -3.4%

NAP1 8.01 3.3% 4.7% -0.1% -0.3% -2.8%NAP2 8.14 0.4% 0.8% 0.0% -0.1% -1.3%NAP3 8.19 -0.5% -0.5% -0.6% -0.9% -3.5%

AC1 8.52 -8.2% -12.0% -3.6% -1.1% -6.8%AC2 7.85 -1.3% -3.7% 0.2% -0.8% -1.9%AC3 7.76 -0.2% -0.3% -0.1% -0.2% -3.0%

Notes: The table reports statistics on the percentage changes to the posted prices charged at locations where pricesare observed (11 residential neighborhoods and 4 commercial districts) under the various counterfactual exercises,computed in the third time period (November 2008). CD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neighborhoods (see text). The first and second row report average andmedian values computed over all 15 neighborhoods with valid prices, respectively. The third and fourth rows reportmedian values computed over the two main commercial districts CD1 and CD2, and over all residential neighbor-hoods, respectively. The first column pertains to the posted prices (NIS) observed in the data. The second columnreports price changes under the counterfactual that reduces the distance disutility parameter β by 50%, and thethird column pertains to reducing, in addition, the “shopping at home” utility parameter κ by 50%. The fourth(fifth) column pertains to improving the amenities at the commercial districts CD1 (CD1 and CD2), while the lastcolumn refers to the counterfactual that adds a supermarket to each residential neighborhood.

32

Table 10: Counterfactual changes to the Average Price Paid (APP)

Retail location Observed APP Reduced travel disutility Improved amenities Additional entry

(NIS) Distance Distance & κ CD1 CD1-CD2

Median residential 7.72 -1.8% -3.2% -4.7% -5.6% -1.6%

NAP1 7.86 0.4% 0.0% -2.2% -3.4% -2.6%NAP2 7.85 -3.5% -5.5% -6.6% -7.3% -0.7%NAP3 7.72 -3.3% -4.7% -7.0% -7.3% -1.2%

AC1 7.98 -5.7% -7.3% -8.6% -8.8% -3.2%AC2 7.67 -2.9% -3.4% -4.7% -6.2% -0.5%AC3 7.28 -1.1% -1.2% -4.1% -4.2% -0.3%

Notes: The table reports the percentage changes in the Average Price Paid (i.e., in the weighted average ofprices that takes into account the probabilities with which residents shop in different destinations) faced byresidents of the 11 residential neighborhoods where prices are observed. All computations performed for thethird time period (November 2008). The first column shows the APP (NIS) corresponding to the observedequilibrium, whereas additional columns show the counterfactual change to the APP given alternative param-eter values: the second column pertains to the counterfactual that reduces the distance disutility parameterβ by 50%, and the third column pertains to reducing, in addition, the “shopping at home” utility parameterκ by 50%. The fourth (fifth) column pertains to improving the amenities at the commercial districts CD1(CD1 and CD2), while the last column refers to the counterfactual that adds a supermarket to each resi-dential neighborhood. CD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, andAffluent Central neighborhoods (see text).

33

Figure 1: Neighborhoods included in the analysis and price levels

34

NAP1

NAP2NAP3

AC3

AC2

AC1

CD1

CD2

77.

58

8.5

9Pr

ice

of c

ompo

site

goo

d (N

IS)

10 15 20Housing prices ('000s NIS per meter squared)

Residential neighborhood Commercial districtFitted values

Note: CD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neighborhoods

Figure 2: Composite good prices plotted against housing prices, November 2008

35

CD1

NAP1 NAP2

NAP3

CD2

AC3

AC2

AC1

77.

58

8.5

Aver

age

Pric

e Pa

id (N

IS)

10 15 20Housing prices ('000s NIS per meter squared)

Average Price Paid (APP) Fitted values

Note: CD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neighborhoods

Figure 3: Average Prices Paid (APP) plotted against housing prices, November 2008

36

CD1

AC3

AC1

NAP3AC2

CD2

NAP2 NAP1

77.

58

8.5

Aver

age

Pric

e Pa

id (N

IS)

0 5000 10000 15000Distance to CD1 (in meters)

Average Price Paid (APP) Fitted values

Note: CD, NAP and AC correspond to Commercial Districts, Non-Affluent Peripheral, and Affluent Central neighborhoods

Figure 4: Average Prices Paid (APP) plotted against distance to CD1, November 2008

37

NAP1

NAP2

AC1

AC2

NAP3AC3

CD2

CD1

Kilometers

0 5

Ratio of probabilities4 - 5.22 - 41.2 - 2No data

Figure 5: Ratio of counterfactual (improved amenities at CD1) to observed probability ofshopping at the main commercial district CD1, November 2008

38

A Online Appendix: Robustness of demand estimates

to the computation of the composite good price

As explained in Section 2.2, we perform robustness checks to verify that our results are not

driven by the way we computed the price for the composite good. Estimation results appear

in Table A1. Elasticities are reported in Table A2.

First, we add locations having at least 9 prices out of the 27 prices for the 27 products.

This increases the number of destinations from 15 to 20 in the first period and 19 in the second

and third periods and the number of observations used in the regression to 2,354. Doing this

decreases the price coefficient and the coefficient of its interaction with housing prices at origin,

although they are still both significant (column 2). This attenuation of the estimates could

reflect increased measurement error in prices brought about by the inclusion of locations with

a different composition of the composite good. This attenuation translates into a decrease in

own prices elasticities from a median elasticity of 4.95 to a median price elasticity of 3.18 (see

Table A2). Remarkably, the estimates of the parameters related to distance remain basically

unchanged. This will also hold for the other robustness checks.

A second check is to use our socioeconomic data to impute prices of products in locations

where they are missing. For each subquarter we compute the mean price (over stores) for

each product and period. We then regress each of these (mean) prices separately on a set

of socioeconomic variables at the neighborhood level, and compute predicted prices for each

product and location.42 In neighborhoods where prices of some products are missing we impute

the predicted prices, and proceed as before to compute the price of the composite good for each

of the destinations where some price data were available.43 The price of the composite good

is now a weighted average of all 27 products. Over all products and locations, the fraction

of imputed prices is 31.5 percent. The imputation procedure generates higher mean prices of

the composite good compared to the observed ones. But these differences are not statistically

significant at the 5 percent significance level. In fact, the top half of the distribution of imputed

42The socioeconomic variables used to predict prices are a subset of the following: number of family house-holds, median age, percentage of married people aged 15 and over, average number of persons per household,percentage of households with 7+ persons in the household, percentage of households with 5+ children up toage 17 in the household, dependency ratio, percentage of those aged 15 and over in the annual civilian laborforce, percentage of those aged 15 and over who did not work in 2008, percentage of Jews born abroad whoimmigrated in 1990-2001, percentage of households residing in self-owned dwellings, percentage of Jews whoseorigin is Israel, percentage of Jews whose continents of origin are America and Oceania, percentage of Jewswhose continent of origin is Europe, percentage of those aged 15 and over with up to 8 years of schooling,percentage of those aged 15 and over with 9-12 years of schooling, percentage of those aged 15 and over with13-15 years of schooling, percentage of those aged 15 and over with 16 or more years of schooling. In addition,we added an indicator for a commercial district and period dummies. The R2′s of these 27 regressions are quitehigh, ranging from 0.45 to 0.93 with a median value of 0.70.

43In 16 observations with missing prices where the imputed price was negative it was substituted for by theminimum imputed price for each product. In neighborhoods that were not sampled in the three periods weimputed prices only for the periods for which we had some price data (these are the neighborhoods with zeronumber of sampled stores in Table D3). Thus, for example, in November 2008 we imputed prices for 23 out ofthe 26 neighborhoods.

39

prices dominates the top half of the distribution of observed prices implying a higher mean price

and variance.

The estimated parameters are somewhat lower than in the baseline specification, again

possibly consistent with attenuation bias due to the measurement error in prices brought about

by the imputation exercise. The estimated own price elasticities are a bit smaller and more

dispersed than in the baseline specification.

In a third robustness check, we estimate the baseline regression using fruits and vegetables

only (11 items).44 The estimated price elasticity is now about a half than in the baseline

specification. This is not surprising since demand for fruits and vegetables is likely to be less

price sensitive than for other products. Note, however, that the sensitivity to distance is about

the same as for the full composite good. We also substitute a very small number (1 NIS) when

expenditures are zero. We can now use the 2070 (46× 15× 3) observations. Results appear in

column (5) of Table A1 and are a bit larger than in the baseline specification. The corresponding

elasticities are shown in Table A2 and are somewhat larger than in the baseline case but, again,

within the same order of magnitude. In a final check we use only price data from supermarkets

and we find that estimated coefficients (column 6 of Table A1) and elasticities are very similar

to the baseline results.

We also estimate a version of our demand model with CPI weights that vary by socioeco-

nomic standing, provided by the CBS. We thus assign differential weights to different origin

neighborhoods. The CBS does not compute expenditure weights for different neighborhoods

but it does compute weights by income level. Specifically, they compute expenditure weights

for very detailed categories of expenditures (but not at the item level as we use in the paper)

by income quintile. In addition, there is a socioeconomic ranking of statistical areas in Israel

and we used this information to assign each of the 46 neighborhoods in Jerusalem to one of

three socio-economic groups: low, middle and high.

We then used the expenditures weights for the first income quintile to compute the price

index faced by residents in neighborhoods in the lowest socio-economic group, the weights of

the third quintile for those in the middle group, and the weights of the fifth quintile for residents

in neighborhoods in the highest socio-economic group. We therefore allow residents of different

(by socio-economic ranking) neighborhoods to face different prices of the composite good even

if they buy in the same destination. The simple correlation coefficient between the original

composite good price and the price computed using income-varying weights is 0.85. Table A3

presents the demand estimates obtained using this approach, with the baseline estimates from

column 6 of Table 6 in the first column.

44In a few locations, the basket is composed of nine or ten fruits and vegetables.

40

Table A1: Robustness results

Variable (1) (2) (3) (4) (5) (6)

Baseline No. of products Imputed Fruits & Including Supermarkets(Col 6 Table 6) in composite >= 9 prices Vegetables Zero exp. only

ln (price at destination) 4.727 3.090 4.107 1.75 5.349 4.061(1.304) (1.200) (1.763) (0.458) (1.766) (1.344)

ln (price) X housing prices -0.232 -0.157 -0.176 -0.077 -0.219 -0.216(.078) (0.064) (0.127) (0.034) (0.132) (.08)

Distance to destination 0.423 0.484 0.452 0.48 0.377 0.409(.12) (0.097) (0.090) (0.103) (0.170) (.13)

Distance X senior citizen 0.004 0.004 0.004 0.005 0.004 0.004(.007) (0.006) (0.005) (0.007) (0.012) (.008)

Distance X driving to work -0.003 -0.004 -0.003 -0.004 0 -0.003(.002) (0.001) (0.001) (0.001) (0.002) (.002)

Shopping at home 1.890 1.873 1.849 1.897 2.16 1.932(.426) (0.294) (0.259) (0.297) (0.485) (.438)

# observations 1819 2354 2968 2091 2070 1633R2 0.784 0.767 0.769 0.757 0.704 0.776

Table A2: Robustness: distribution of estimated elasticities (absolute value)

Own price elasticity

Specification mean sd min p10 p25 p50 p75 p90 max NBaseline (col 6 Table 6) σ = 0.7 4.82 0.92 3.00 3.86 3.99 4.95 5.87 5.95 6.13 15Baseline (col 6 Table 6) σ = 0.8 6.43 1.37 3.78 5.01 5.31 6.54 7.94 8.32 8.47 15

Composite with 9 or more products 3.08 0.77 1.67 1.91 2.51 3.18 3.54 4.12 4.21 19Imputed prices 4.40 1.26 2.30 2.67 3.01 4.52 5.34 5.89 6.34 23Fruits and Vegetables 2.51 0.49 1.55 1.68 2.23 2.58 2.84 3.20 3.22 19Including zero Exp. 6.60 0.96 4.75 5.55 5.68 6.59 7.29 8.02 8.16 15Supermarkets only 3.84 0.87 2.15 2.94 3.09 3.88 4.75 4.96 5.20 15

Distance semi-elasticity

Specification mean sd min p10 p25 p50 p75 p90 max NBaseline (col 6 Table 6) 0.35 0.06 0.06 0.28 0.31 0.35 0.39 0.42 0.45 690

Composite with 9 or more products 0.37 0.07 0.06 0.29 0.33 0.37 0.43 0.48 0.50 874Imputed prices 0.37 0.05 0.16 0.31 0.33 0.37 0.42 0.45 0.47 1,058Fruits and Vegetables 0.37 0.07 0.13 0.28 0.32 0.37 0.44 0.48 0.50 798Including zero Exp. 0.40 0.05 0.09 0.36 0.39 0.40 0.42 0.44 0.48 690Supermarkets only 0.34 0.06 0.06 0.27 0.30 0.34 0.38 0.41 0.44 645

Notes: Elasticities are computed for November 2008. σ = 0.7 is used except in row 2 of top panel. Priceelasticities are computed for each destination. Prices were imputed for 23 out of the 26 neighborhoods inNovember 2008. Distance semi-elasticities are computed for each origin-destination pair (e.g., 46x15=690).

41

Table A3: Demand estimates with income-varying weights

(1) (2)

Variable Baseline (Col 6 from Table 6) Using income-varying weights

ln (price at destination) 4.727 5.138(1.304) (1.940)

ln (price at destination) X housing prices -0.232 -0.450(0.078) (0.104)

Distance to destination 0.423 0.422(0.120) (0.119)

Distance to destination X senior citizen 0.004 0.004(0.007) (0.007)

Distance to destination X driving to work -0.003 -0.003(0.002) (0.002)

Shopping at home 1.890 1.888(0.426) (0.423)

# observations 1819 1819R2 (0.784) (0.784)

42

B Online Appendix: Counterfactual analyses for σ = 0.8

Table B1: Counterfactual changes to posted prices, σ = 0.8

Retail location Observed price Reduced travel disutility Improved amenities Additional entry

Distance Distance & κ CD1 CD1-CD2

Average (all) 7.80 -0.7% -1.1% -0.2% -0.2% -2.3%

Median (all) 7.85 -0.9% -2.1% 0.0% -0.5% -3.7%

Median CD1-CD2 6.98 -0.5% -0.5% 0.2% 0.1% 0.0%

Median residential 7.87 -1.2% -1.8% -0.1% -0.5% -1.9%

NAP1 8.01 2.6% 3.7% 0.0% -0.2% -2.4%NAP2 8.14 0.3% 0.7% 0.1% 0.0% -1.1%NAP3 8.19 -0.3% -0.2% -0.4% -0.5% -3.1%

AC1 8.52 -6.3% -9.3% -2.7% -0.8% -5.9%AC2 7.85 -0.9% -2.7% 0.2% -0.5% -1.6%AC3 7.76 -0.2% -0.2% -0.1% -0.1% -2.7%

Notes: The table reports the corresponding values to those in Table 9, but using a value σ = 0.8 rather thanσ = 0.7. See notes to table 9 for additional details.

Table B2: Counterfactual changes to the Average Price Paid (APP), σ = 0.8

Retail location Observed price Reduced travel disutility Improved amenities Additional entry

Distance Distance & κ CD1 CD1-CD2

Median residential 7.72 -1.6% -3.0% -4.8% -5.7% -1.4%

NAP1 7.86 0.6% 0.3% -2.1% -3.3% -2.1%NAP2 7.85 -3.4% -5.4% -6.8% -7.3% -0.6%NAP3 7.72 -3.1% -4.5% -7.1% -7.4% -1.1%

AC1 7.98 -4.9% -6.7% -8.6% -8.8% -2.8%AC2 7.67 -2.7% -3.1% -4.8% -6.4% -0.4%AC3 7.28 -0.9% -1.0% -4.3% -4.3% -0.3%

Notes: The table reports the corresponding values to those in Table 10, but using a value σ = 0.8 ratherthan σ = 0.7. See notes to table 10 for additional details.

43

C Online appendix: Neighborhoods, subquarters and

demographics

While distinct neighborhoods with established identities are a key feature of Jerusalem, there

is no formal statistical definition that precisely matches the notion of a “neighborhood.” We

therefore use the Central Bureau of Statistics’s (CBS) closely-related concept of a subquarter.

A subquarter includes several territorially-contiguous statistical areas.45 We use the terms

“neighborhood” and “subquarter” interchangeably.

We defined the six commercial districts (appearing in bold in Table C1 below) as collections

of statistical areas that are predominantly commercial with minimal residential presence. These

areas were typically carved out of a larger subquarter that was partitioned into primarily resi-

dential, and primarily non-residential collections of statistical areas. The two major commercial

districts are Talpiot and Givat Shaul denoted by CD1 and CD2 in the text.

Thus, neighborhoods are identified with the subquarters defined by the CBS with some

exceptions: 1) the commercial districts that were carved out from existing subquarters as

mentioned above, and 2) four subquarters that were added to accommodate the expenditure

data received from the credit card company. These additional subquarters share some of the

statistical areas with other subquarters and are denoted in Table C1 with a star *. Although

these four subquarters share the same statistical areas (and therefore the same demographics)

they do have different zipcodes and therefore different expenditure data.

Table C1 presents our 46 subquarters (neighborhoods) and provides the statistical areas

that are included in each neighborhood. Tables C2-C3 provide neighborhood-level statistics on

demographics and distances.

45A statistical area is a small geographic unit as homogeneous as possible, generally including 3,000 — 4,000persons in residential areas. http://www.cbs.gov.il/mifkad/mifkad 2008/hagdarot e.pdf.

44

Table C1: Composition of residential and commercial neighborhoods

Subquarter (neighborhood) statistical areas

Neve Yaaqov 111 112 113 114 115 116Pisgat Zeev North 121 122 123 124 125Pisgat Zeev East 131 132 133 134 135 136Pisgat Zeev (North - West & West) * 135 136Ramat Shlomo 411 412 413Ramot Allon North 421 422 423 424 425 426Ramot Allon 431 432 433 434 435 436Ramot Allon South * 435Har Hahozvim, Sanhedria 511 512 513 514 515Ramat Eshkol, Givat-Mivtar 521 522 523Maalot Dafna, Shmuel Hanavi 531 532 533Givat Shapira 541 542 543Mamila, Morasha 811 812Geula, Mea Shearim 821 822 823 824 825 826Makor Baruch, Zichron Moshe 831 832 833 834 835 836City Center 841 842 843 844 845 846 847Nahlaot, Zichronot 851 852 854 855 856 857 858Rehavya 861 862 863 864Romema 911 912 913 915 916Givat Shaul 921 922 923 925Har Nof 931 932 933 934Qiryat Moshe, Bet HaKerem 1011 1012 1013 1014 1015 1016Nayot 1021 1022 1023 1024Bayit VaGan 1031 1032 1033 1034 1035Ramat Sharet, Ramat Denya 1041 1042 1043 1044Qiryat HaYovel North 1121 1122 1123 1124Qiryat HaYovel South 1131 1132 1133 1134Qiryat Menahem, Ir Gannim 1141 1142 1143 1144 1145 1146 1147Manahat slopes * 1147Gonen (Qatamon) 1211 1212 1213 1214 1215 1216 1217Rassco, Givat Mordekhay 1221 1222 1223German Colony, Gonen (Old Qatamon) 1311 1312 1313 1314Qomemiyyut (Talbiya), YMCA Compound 1321 1322Baqa, Abu Tor, Yemin Moshe 1331 1332 1333 1334 1335 1336Talpiot, Arnona, Mekor Haym 1341 1342 1343 1344 1346East Talpiot 1351 1352 1353 1354 1355East Talpiot (East) * 1355Homat Shmuel (Har Homa) 1621 1622 1623Gilo East 1631 1632 1633 1634Gilo West 1641 1642 1643 1644Talpiot CD 1345 Talpiot - Industrial & Commercial Area,

Yad Haruzim st.Givat Shaul CD 924 Givat Shaul Industrial Area and ”B”,

Menuhot Cemetery, Kanfei NesharimMalcha CD 1146 Tedy Stadium, Biblical Zoo, Jerusalem MallRomema CD 914 Romema, Industrial Area, Etz Haim,Central Bus Station CDMahane Yehuda CD 853 Beit Yaakov, Clal Ctr., Mahane Yehuda Market

Notes: The table presents our 46 subquarters (neighborhoods), and provides the statisticalareas that are included in each neighborhood. For residential neighborhoods, the statisticalareas included follow the CBS definitions. For commercial districts (in bold), the includedstatistical areas were determined by the authors and their explicit names are provided. Res-idential neighborhoods marked with an * mean that the neighborhood shares portions of thesame statistical areas with preceding neighborhood. A common statistical area was dividedinto two subquarters according to the zipcodes of the expenditure data.

45

Table C2: Demographics, housing prices and number of supermarkets

Population Household Housing % driving % car % senior No. ofNeighborhood (000s) size price to work ownership citizens supermarkets

Neve Yaaqov 18.3 3.9 9.5 21.2 28.6 7.6 1Pisgat Zeev North 17.7 3.3 8.8 48.3 66.5 10.4 1Pisgat Zeev East 21.7 3.6 9.7 59.2 73.5 7.6 0Pisgat Zeev (No.West & West) 21.7 3.6 9.2 59.2 73.5 7.6 0Ramat Shlomo 14.1 6.1 12.2 23.8 35 1.1 0Ramot Allon North 23.1 4.9 11.9 32.7 39.9 2.5 1Ramot Allon 16.6 4.1 12.2 51.4 61.3 5.6 0Ramot Allon South 16.6 4.1 12.0 51.4 61.3 5.6 0Har Hahozvim, Sanhedria 15.8 5.3 15.7 9.9 14.7 4.6 0Ramat Eshkol, Givat-Mivtar 10.2 3.9 15.2 27.5 34.4 12.1 0Maalot Dafna, Shmuel Hanavi 8.7 4 13.3 17.1 21.8 7 0Givat Shapira 9.3 2.3 10.7 56.3 65.9 10.6 2Mamila, Morasha 13 3.3 15.6 9.9 12.4 10.7 0Geula, Mea Shearim 28.7 4.6 13.9 7.5 6.9 5.9 0Makor Baruch, Zichron Moshe 13 3.3 13.2 9.9 12.4 10.7 0City Center 6.2 1.9 13.7 13.6 24 15.4 2Nahlaot, Zichronot 9.1 2.1 15.5 27.4 35.7 12.5 0Rehavya 7.5 2 21.1 42.5 57.6 25.6 1Romema 21.1 4.5 15.8 11.4 10.7 7.5 1Givat Shaul 10.5 4.2 13.0 33.8 40.6 7 0Har Nof 15.8 4.3 13.8 36.1 49.2 6.4 1Qiryat Moshe, Bet HaKerem 23.3 2.7 15.8 49.8 62.4 16.7 2Nayot 23.3 2.7 15.1 49.8 62.4 16.7 1Bayit VaGan 18.1 3.4 15.9 30.7 39.1 12.3 0Ramat Sharet, Ramat Denya 8.5 3.3 14.9 68.1 85.4 8.9 0Qiryat HaYovel North 10.6 2.7 11.9 46 54.6 16.9 0Qiryat HaYovel South 10.6 2.4 11.5 44.8 49.4 16.3 1Qiryat Menahem, Ir Gannim 17.5 3.3 11.8 57 62.5 10.2 1Manahat slopes 17.5 3.3 14.9 57 62.5 10.2 0Gonen (Qatamon) 23.5 2.8 11.7 39.7 50.7 11.9 0Rassco, Givat Mordekhay 13.5 2.4 15.1 51.5 62.9 14.4 1German Colony, Gonen 10 2.5 19.7 52 69.6 16.3 0Qomemiyyut (Talbiya), YMCA 10 2.5 20.7 52 69.6 16.3 0Baqa, Abu Tor, Yemin Moshe 11 2.9 15.0 51.7 67 16.4 1Talpiot, Arnona, Mekor Haim 13.8 2.8 13.6 55.5 67.9 18 0East Talpiot 13.9 2.9 9.5 55.3 60.8 9.5 0East Talpiot (East) 13.9 2.9 9.5 55.3 60.8 9.5 0Homat Shmuel (Har Homa) 9.8 4 10.4 66.7 89.3 2.3 0Gilo East 18.7 3.1 9.4 53.2 65.5 11.6 0Gilo West 10.4 3.4 9.3 63.7 77.6 8.9 0Talpiot CD 11 2.9 9.5 51.7 67 16.4 5Givat Shaul CD 10.5 4.2 13.0 33.8 40.6 7 3Malcha CD 17.5 3.3 14.9 57 62.5 10.2 1Romema CD 21.1 4.5 15.8 11.4 10.7 7.5 3Central Bus Station CD 21.1 4.5 15.8 11.4 10.7 7.5 0Mahane Yehuda CD 13 3.3 13.2 9.9 12.4 10.7 1

Notes: Commercial districts have associated demographics because they also contain a small residential neigh-borhood. Housing prices = the 2007-2008 average price per square meter in thousands of dollars. Driving towork = percentage of those aged 15 and over who used a private car or a commercial vehicle (as a driver) astheir main means of getting to work in the determinant week. Car ownership = percentage of households us-ing at least one car. Senior citizens = percentage of individuals above age 65. Source: CBS. The number ofsupermarkets includes all supermarkets in the neighborhood, not just those where prices were sampled.

46

Table C3: Distances (in km)

Neighborhood Distance to:

All neighborhoods City Commercial Districts(mean) center CD 1 CD 2

Neve Yaaqov 10.8 9.2 13.2 12.0Pisgat Zeev North 9.3 7.5 11.6 10.6Pisgat Zeev East 8.9 7.0 11.0 10.2Pisgat Zeev (North - West & West) 8.1 6.1 10.2 9.4Ramat Shlomo 7.0 5.1 9.4 6.9Ramot Allon North 7.7 6.5 10.6 7.0Ramot Allon 7.3 6.0 10.0 6.1Ramot Allon South 7.3 6.1 10.2 6.6Har Hahozvim, Sanhedria 4.9 2.4 6.7 4.6Ramat Eshkol, Givat-Mivtar 5.5 3.0 7.2 5.7Maalot Dafna, Shmuel Hanavi 4.9 2.0 6.1 5.1Givat Shapira 6.4 3.7 7.8 7.1Mamila, Morasha 4.6 0.9 4.3 5.1Geula, Mea Shearim 4.5 1.2 5.5 4.5Makor Baruch, Zichron Moshe 4.4 1.3 5.4 3.7City Center 4.4 0.6 4.4 4.4Nahlaot, Zichronot 4.3 1.1 4.5 3.7Rehavya 4.4 1.5 3.6 4.5Romema 5.0 3.0 6.6 3.4Givat Shaul 5.8 4.1 7.5 2.8Har Nof 6.6 5.1 8.1 2.8Qiryat Moshe, Bet HaKerem 4.8 3.5 5.5 2.6Nayot 4.8 2.9 4.6 3.8Bayit VaGan 6.0 5.7 5.7 4.7Ramat Sharet, Ramat Denya 6.5 6.5 4.8 5.9Qiryat HaYovel North 6.1 6.1 5.4 5.0Qiryat HaYovel South 6.5 6.6 5.0 5.9Qiryat Menahem, Ir Gannim 8.3 8.5 7.0 7.6Manahat slopes 6.0 5.6 3.6 6.5Gonen (Qatamon) 5.2 4.0 1.9 6.1Rassco, Givat Mordekhay 4.8 3.0 2.8 5.0German Colony, Gonen (Old Qatamon) 4.7 2.5 2.3 5.6Qomemiyyut (Talbiya), YMCA Compound 4.5 1.3 3.4 5.2Baqa, Abu Tor, Yemin Moshe 5.2 2.8 2.1 6.5Talpiot, Arnona, Mekor Haim 5.7 4.0 1.2 7.5East Talpiot 6.9 5.0 3.0 8.8East Talpiot (East) 6.9 4.9 3.3 8.8Homat Shmuel (Har Homa) 8.3 7.2 3.4 10.4Gilo East 7.6 7.2 3.6 9.0Gilo West 8.8 8.4 4.9 10.2Talpiot (CD 1) 5.7 4.4 0.0 7.5Givat Shaul (CD 2) 6.0 4.4 7.5 0.0Malcha CD 5.7 5.2 3.1 6.2Romema CD 4.5 2.0 5.6 3.1Central Bus Station CD 4.5 2.0 5.6 3.1Mahane Yehuda CD 4.2 1.1 5.0 3.5Average 6.1 4.3 5.7 6.0Standard deviation 1.6 2.3 2.9 2.5Median 5.8 4.3 5.4 5.8

Notes: Distances in kilometers between each neighborhood and 1) the city center, 2) the two promi-nent commercial centers CD1 and CD2, and 3) all other neighborhoods (mean distance). Source:CBS.

47

D Online appendix: Products, prices and expenditures

Table D1: Definition of products

1 Waffles simple packed waffles, non-coated,same brand2 Mayonnaise low-fat mayonnaise, same brand3 Cottage cheese 250 gr container of same brand4 Sugar packed sugar, same brand, 1kg5 Chocolate bar regular milk chocolate, same brand6 Mineral water in plastic bottle, 1.5 liter7 Coca cola in plastic bottle, 1.5 liter8 Ketchup same brand9 Tea regualr tea, teabags, same brand10 Turkish coffee packaged roasted and ground turkish coffee, same brand11 Cocoa powder instant chocolate powder, same brand12 Green peas (can) garden variety, same brand13 Hummus (salad) hummus salad, not fresh, same brand14 Cucumbers fresh standard cucumbers, type A, 1kg15 Onion dry onion, type A, 1kg16 Carrots medium size fresh carrots, type A, 1kg17 Eggplants medium size fresh eggplants, type A, 1kg18 Cabbage (white) white fresh cabbage, 1kg19 Cauliflower fresh cauliflower, type A, 1kg20 Potatoes fresh potatoes, type A, 1kg21 Tomatoes round tomatoes, type A, 1kg22 Apples granny smith apples, type A, 1kg23 Bananas type A, 1 kg24 Lemons fresh, type A, 1kg25 Fabric softener same brand26 Dishwasher detergent in plastic bottle, same brand27 Shaving cream/gel same brand

48

Tab

leD

2:P

roduct

-sp

ecifi

cpri

cedis

trib

uti

ons

(NIS

)

Pro

du

ctM

ean

Coeffi

cien

t#

stor

esP

rod

uct

Mea

nC

oeffi

cien

t#

store

sP

rod

uct

Mea

nC

oeffi

cien

t#

store

sp

rice

ofV

aria

tion

pri

ceof

Vari

ati

on

pri

ceof

Vari

ati

on

Waffl

es

Tu

rkis

hcoff

ee

Cau

lifl

ow

er

Sep

-07

10.4

0.14

24S

ep-0

75.8

0.0

923

Sep

-07

7.3

0.3

225

Nov

-07

10.2

0.18

22N

ov-0

75.7

0.1

123

Nov

-07

5.9

0.1

922

Nov

-08

11.1

0.24

20N

ov-0

87

0.0

723

Nov

-08

6.6

0.2

423

Mayon

nais

eC

ocoa

pow

der

Pota

toes

Sep

-07

7.6

0.12

22S

ep-0

710.3

0.1

223

Sep

-07

40.2

337

Nov

-07

90.

2121

Nov

-07

10.5

0.1

223

Nov

-07

4.2

0.2

637

Nov

-08

9.6

0.14

16N

ov-0

810.7

0.1

122

Nov

-08

4.8

0.2

535

Cott

age

cheese

Gre

en

peas

(can

)T

om

ato

es

Sep

-07

5.3

0.04

23S

ep-0

75.2

0.1

016

Sep

-07

6.1

0.3

337

Nov

-07

5.8

0.03

25N

ov-0

75.2

0.1

016

Nov

-07

50.3

437

Nov

-08

60.

0522

Nov

-08

5.9

0.1

214

Nov

-08

6.9

0.3

335

Su

gar

Hu

mm

us

(sala

d)

Ap

ple

sS

ep-0

73.

60.

2224

Sep

-07

90.1

117

Sep

-07

90.2

036

Nov

-07

3.6

0.22

23N

ov-0

79.2

0.0

518

Nov

-07

9.1

0.1

234

Nov

-08

3.4

0.26

24N

ov-0

810.6

0.1

014

Nov

-08

9.6

0.1

833

Ch

ocola

teb

ar

Cu

cu

mb

ers

Ban

an

as

Sep

-07

4.4

0.11

23S

ep-0

74.6

0.2

837

Sep

-07

6.3

0.1

335

Nov

-07

4.5

0.11

23N

ov-0

75.8

0.1

737

Nov

-07

5.6

0.3

035

Nov

-08

5.1

0.12

23N

ov-0

84.8

0.2

935

Nov

-08

7.8

0.2

333

Min

era

lw

ate

rO

nio

nL

em

on

sS

ep-0

712

.80.

1121

Sep

-07

2.8

0.3

237

Sep

-07

11.7

0.2

238

Nov

-07

12.7

0.15

20N

ov-0

73.2

0.3

436

Nov

-07

8.1

0.2

536

Nov

-08

12.3

0.28

20N

ov-0

83.7

0.3

535

Nov

-08

10.4

0.3

735

Coca

cola

Carr

ots

Fab

ric

s.S

ep-0

75.

50.

1825

Sep

-07

4.9

0.1

837

Sep

-07

20.8

0.0

821

Nov

-07

5.5

0.18

25N

ov-0

75.1

0.1

836

Nov

-07

19.9

0.1

625

Nov

-08

5.9

0.17

24N

ov-0

85.6

0.3

832

Nov

-08

22.1

0.0

722

Ketc

hu

pE

ggp

lants

Dis

hw

ash

er

d.

Sep

-07

11.1

0.14

24S

ep-0

74

0.4

038

Sep

-07

10.8

0.1

216

Nov

-07

10.9

0.14

24N

ov-0

73.7

0.4

135

Nov

-07

11.9

0.1

019

Nov

-08

110.

1523

Nov

-08

4.7

0.3

433

Nov

-08

11.1

0.2

023

Tea

Cab

bage

(wh

ite)

Sh

avin

gc/g

Sep

-07

15.8

0.15

22S

ep-0

74.7

0.5

133

Sep

-07

22.1

0.2

022

Nov

-07

16.2

0.15

23N

ov-0

73.7

0.5

732

Nov

-07

23.2

0.2

216

Nov

-08

17.1

0.15

20N

ov-0

85.1

0.6

131

Nov

-08

23.5

0.1

618

49

Table D3: Number of sampled stores and of observed products

# sampled stores # observed products # supermarketsNeigborhood Sep2007 Nov2007 Nov2008 Sep2007 Nov2007 Nov2008

Neve Yaaqov 1 1 1 27 27 27 1Pisgat Zeev North 1 1 1 26 26 27 1Ramot Allon North 2 2 2 24 25 25 1Ramat Eshkol, G. Mivtar 1 1 1 11 10 9 0M. Dafna, S. Hanavi 1 0 0 10 0 0 0Givat Shapira 2 2 2 27 27 27 2Geula, Mea Shearim 3 4 3 12 12 13 0City Center 1 2 2 6 7 6 2Rehavya 2 2 2 24 25 24 1Romema 2 2 2 24 23 22 1Givat Shaul 1 1 1 3 4 3 0Har Nof 1 1 1 25 21 22 1Qiryat Moshe, B. Hakerem 3 3 3 27 27 27 2Nayot 1 1 1 11 11 11 1Ramat Sharet-Denya 1 1 0 1 1 0 0Qiryat HaYovel South 3 2 2 27 26 26 1Rassco, Givat Mordekhay 2 2 2 26 27 27 1Baqa, Abu Tor, Y. Moshe 1 1 1 26 25 23 1Talpiot, Arnona, M. Haim 1 1 1 4 4 2 0Gilo East 0 1 0 0 1 0 0Gilo West 2 2 2 12 13 12 0Talpiot CD 7 7 7 27 27 27 5Givat Shaul CD 3 3 3 27 27 26 3Malcha CD 1 1 1 3 4 4 1Romema CD 1 1 1 27 27 23 3Mahane Yehuda CD 10 10 9 25 24 24 1

Notes: The 15 neighborhoods with price data for at least 21 out of the 27 products appear in bold.

50

Table D4: Product composition of composite good

Neighborhood

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15ProductWaffles 3 3 3 3 3 1 3 3 3 3 3 3 3 3 1Mayonnaise 3 3 3 3 2 2 1 3 3 3 3 3 3 2 3Cottage ch. 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Sugar 3 3 3 3 3 2 3 3 3 3 3 3 3 3 3Chocolate bar 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Mineral water 3 2 1 3 3 3 3 3 3 3 1 3 3 3 3Coca cola 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Ketchup 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Tea 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Turkish coffee 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Cocoa powder 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Potatoes 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Tomatoes 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Cucumbers 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Onion 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Carrots 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Eggplants 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Cabbage 3 3 3 3 3 2 3 3 3 3 3 3 3 3 3Cauliflower 3 3 3 3 3 0 0 3 3 3 1 3 3 2 3Apples 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Bananas 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Lemons 3 3 3 3 3 3 2 3 3 3 3 3 3 3 3G. peas (can) 3 3 2 3 1 2 0 3 3 3 2 3 3 2 2Hummus 3 3 3 3 1 2 1 3 1 3 2 3 2 3 3Fabric soft. 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3Dishwasher d. 3 2 2 3 1 3 3 3 3 2 3 3 3 3 1Shaving c/g 3 3 0 3 2 1 1 3 3 3 2 3 3 2 0

Notes: Entries are the number of times a product (row) appears in a neighbor-hood (column) over the three periods. A ”3” means that the products was alwaysin the composite basket, while a ”0” means that it was never included in the bas-ket. In both cases, there is no change in the composition of the basket over time.The 15 neighborhoods are: 1= Neve Yaaqov, 2= Pisgat Zeev N., 3=Ramot AlonN., 4=givat Shapira, 5=Rehavia, 6=Romema, 7=Har Nof, 8=Qiryat Moshe, BetHakerem. 9=Qiryat Hayovel South, 10=Rasko, Givat Mordekhay, 11=Baqa, AbuTor, Yemin Moshe, 12= Talpiot (CD1), 13=Givat Shaul (CD2), 14=Romema CD,15=mahane Yehuda CD.

51

Table D5: Distribution of products across neighborhoods

Sep-07 Nov-07 Nov-08

Waffles 15 13 13Low fat mayonnaise 15 14 11Cottage cheese 15 15 15Sugar 15 14 15Chocolate bar 15 15 15Mineral water 14 12 14Coca cola 15 15 15Ketchup 15 15 15Tea 15 15 15Turkish coffee 15 15 15Cocoa powder 15 15 15Potatoes 15 15 15Tomatoes 15 15 15Cucumbers 15 15 15Onion 15 15 15Carrots 15 15 15Eggplants 15 15 15Cabbage (white) 14 15 15Cauliflower 12 12 12Apples 15 15 15Bananas 15 15 15Lemons 15 14 15Green peas (can) 13 13 9Hummus 13 13 10Fabric softener 15 15 15Dishwasher detergent 10 13 15Shaving cream/gel 13 11 8

Notes: Entries are the number of neighborhoods inwhich a product has non-missing price data per pe-riod.

52

Table D6: Composite good prices (NIS) across Jerusalem neighborhoods

Sep-07 Nov-07 Nov-08

Ramot Allon north 6.23 Talpyiot CD (CD1) 6.15 Talpyiot CD (CD1) 6.89Talpyiot CD (CD1) 6.33 Ramot Allon north 6.56 Givat Shaul CD (CD2) 7.07Mahane Yehuda CD 6.84 Mahane Yehuda CD 6.81 Mahane Yehuda CD 7.20Romema CD 7.03 Pisgat Zeev North 6.89 Pisgat Zeev North 7.36Har Nof 7.13 Har Nof 6.93 Ramot Allon north 7.61Neve Yaaqov 7.15 Romema CD 6.99 Har Nof 7.62Rassco, Givat Mordekhay 7.32 Baqa, Abu Tor, Yemin Moshe 7.06 Baqa, Abu Tor, Yemin Moshe 7.76Pisgat Zeev North 7.34 Rehavya 7.27 Qiryat Moshe, Bet Hakerem 7.85Givat Shaul CD (CD2) 7.45 Givat Shaul CD (CD2) 7.30 Rassco, Givat Mordekhay 7.87Giv’at Shapira 7.54 Neve Yaaqov 7.31 Neve Yaaqov 8.01Qiryat Moshe, Bet Hakerem 7.55 Rassco, Givat Mordekhay 7.34 Giv’at Shapira 8.14Romema 7.61 Qiryat Ha-Yovel south 7.36 Romema 8.17Baqa, Abu Tor, Yemin Moshe 7.68 Romema 7.38 Qiryat Ha-Yovel south 8.19Qiryat Ha-Yovel south 7.80 Giv’at Shapira 7.39 Rehavya 8.52Rehavya 8.01 Qiryat Moshe, Bet Hakerem 7.61 Romema CD 8.69

Mean 7.27 7.09 7.80Standard deviation 0.50 0.38 0.52

Notes: Source: CBS.

53

Table D7: Credit Card Expenditures

Neighborhood Fraction spent at

Own neighborhood CD1 CD2Neve Yaaqov 0.25 0.03 0.02Pisgat Zeev North 0.68 0.10 0.03Pisgat Zeev East 0.22 0.23 0.06Pisgat Zeev (North - West & West) 0.01 0.24 0.08Ramat Shlomo 0.18 0.01 0.02Ramot Allon North 0.25 0.12 0.06Ramot Allon 0.15 0.15 0.08Ramot Allon South 0.31 0.18 0.11Har Hahozvim, Sanhedria 0.08 0.01 0.02Ramat Eshkol, Givat-Mivtar 0.56 0.05 0.02Maalot Dafna, Shmuel Hanavi 0.18 0.08 0.02Givat Shapira 0.42 0.18 0.04Mamila, Morasha 0.05 0.29 0.06Geula, Mea Shearim 0.24 0.06 0.02Makor Baruch, Zichron Moshe 0.03 0.04 0.02City Center 0.10 0.16 0.05Nahlaot, Zichronot 0.03 0.17 0.04Rehavya 0.44 0.19 0.03Romema 0.54 0.03 0.02Givat Shaul 0.60 0.03 0.16Har Nof 0.30 0.01 0.31Qiryat Moshe, Bet HaKerem 0.14 0.16 0.18Nayot 0.08 0.14 0.20Bayit VaGan 0.05 0.17 0.10Ramat Sharet, Ramat Denya 0.12 0.31 0.07Qiryat HaYovel North 0.21 0.21 0.07Qiryat HaYovel South 0.33 0.31 0.05Qiryat Menahem, Ir Gannim 0.52 0.21 0.03Manahat slopes 0.07 0.55 0.06Gonen (Qatamon) 0.07 0.55 0.03Rassco, Givat Mordekhay 0.31 0.47 0.03German Colony, Gonen (Old Qatamon) 0.07 0.61 0.03Qomemiyyut (Talbiya), YMCA Compound 0.01 0.29 0.05Baqa, Abu Tor, Yemin Moshe 0.00 0.65 0.02Talpiot, Arnona, Mekor Haim 0.15 0.71 0.02East Talpiot 0.01 0.71 0.03East Talpiot (East) 0.01 0.66 0.02Homat Shmuel (Har Homa) 0.00 0.72 0.03Gilo East 0.21 0.46 0.02Gilo West 0.26 0.46 0.03Talpiot commercial district 0.76 0.76 0.03Givat Shaul commercial district 0.41 0.06 0.41Malcha commercial district 0.01 0.60 0.05Romema commercial district 0.60 0.04 0.03Central Bus Station 0.14 0.16 0.01Mahane Yehuda 0.06 0.26 0.08

Mean 0.22 0.27 0.06Median 0.16 0.19 0.03

Notes: Entries are expenditure fractions averaged over the three periods ofdata.

54

E Online appendix: Observed and counterfactual posted

prices and Average Prices Paid in all neighborhoods

Table E1 presents the counterfactual price changes in the 15 neighborhoods where prices could

be computed using at least 21 products. Table E2 presents changes in the APP in all 46

neighborhoods.

55

Tab

leE

1:C

ounte

rfac

tual

chan

ges

top

oste

dpri

ces

Nei

ghb

orh

ood

Ob

serv

edp

rice

Red

uce

dtr

avel

dis

uti

lity

Impro

ved

am

enit

ies

Ad

dit

ion

al

entr

y

Dis

tan

ceD

ista

nce

&κ

CD

1C

D1-C

D2

Nev

eY

aaqov

(NA

P1)

8.0

13.3

%4.7

%-0

.1%

-0.3

%-2

.8%

Pis

gat

Zee

vN

orth

7.3

60.3

%0.7

%-0

.9%

-1.2

%-3

.4%

Ram

otA

llon

nor

th7.6

10.2

%0.3

%-0

.4%

-0.5

%-3

.3%

Giv

atS

hap

ira

(NA

P2)

8.1

40.4

%0.8

%0.0

%-0

.1%

-1.3

%R

ehav

ya(A

C1)

8.5

2-8

.2%

-12.0

%-3

.6%

-1.1

%-6

.8%

Rom

ema

8.1

7-1

.3%

-2.7

%1.2

%1.9

%-4

.4%

Har

Nof

7.6

2-0

.7%

-1.7

%0.0

%-1

.3%

-4.3

%Q

irya

tM

osh

e,B

.H

aKer

em(A

C2)

7.8

5-1

.3%

-3.7

%0.2

%-0

.8%

-1.9

%Q

irya

tH

aYov

elS

outh

(NA

P3)

8.1

9-0

.5%

-0.5

%-0

.6%

-0.9

%-3

.5%

Ras

sco,

Giv

atM

ord

ekh

ay7.8

7-1

.7%

-3.4

%-0

.7%

-0.9

%-4

.6%

Baq

a,A

bu

Tor

,Y

.M

osh

e(A

C3)

7.7

6-0

.2%

-0.3

%-0

.1%

-0.2

%-3

.0%

Tal

pio

t(C

D1)

6.8

9-0

.3%

-0.2

%0.5

%0.2

%0.0

%G

ivat

Sh

aul

(CD

2)7.0

7-1

.1%

-1.0

%0.3

%0.3

%0.0

%R

omem

aC

D8.6

9-1

.0%

-1.0

%0.4

%0.3

%0.1

%M

ahan

eY

ehu

da

CD

7.2

0-1

.5%

-1.4

%0.1

%-0

.2%

0.1

%

Mea

n(r

esid

enti

al)

-0.9

%-1

.6%

-0.5

%-0

.5%

-3.6

%M

edia

n(r

esid

enti

al)

-0.5

%-0

.5%

-0.1

%-0

.8%

-3.4

%

Notes:

Th

eta

ble

rep

orts

the

per

centa

gech

an

ges

inp

rice

sch

arg

edin

the

15

nei

ghb

orh

ood

sw

her

eth

eco

mp

osi

tegood

pri

ceco

uld

be

com

pu

ted

usi

ng

atle

ast

21go

od

s.T

he

cou

nte

rfact

uals

,p

erfo

rmed

inth

eth

ird

sam

ple

per

iod

,are

des

crib

edin

the

text

ind

etai

l.

56

Table E2: Counterfactual changes to the Average Price Paid (APP)

Retail location Observed price Reduced travel disutility Improved amenities Additional entry

Distance Distance & κ CD1 CD1-CD2

Neve Yaaqov (NAP1) 7.86 0.4% 0.0% -2.2% -3.4% -2.6%Pisgat Zeev North 7.48 -1.5% -1.5% -3.2% -3.7% -2.7%Pisgat Zeev East 7.67 -4.2% -4.2% -6.3% -6.8% -0.8%Pisgat Zeev (NW & W) 7.46 -3.0% -2.9% -4.6% -4.9% -1.5%Ramat Shlomo 8.20 -1.1% -1.4% -0.5% -3.6% -1.2%Ramot Allon north 7.86 -3.3% -3.5% -5.2% -6.7% -1.6%Ramot Allon 7.83 -3.6% -3.8% -5.5% -6.9% -1.1%Ramot Allon South 7.75 -4.8% -4.9% -6.0% -6.9% -0.7%Har Hahozvim, Sanhedria 8.29 -1.4% -1.7% -0.4% -3.4% -1.4%Ramat Eshkol, Givat-Mivtar 8.12 -2.6% -2.6% -4.2% -5.6% -0.3%Maalot Dafna, Shmuel Hanavi 8.07 -2.8% -2.9% -4.9% -6.1% -0.4%Givat Shapira (NAP2) 7.85 -3.5% -5.5% -6.6% -7.3% -0.7%Mamila, Morasha 7.80 -3.6% -3.9% -7.4% -7.8% -0.7%Geula, Mea Shearim 8.18 -2.2% -2.3% -4.0% -6.0% -0.5%Makor Baruch, Zichron Moshe 8.28 -2.5% -3.1% -2.6% -6.0% -1.1%City Center 7.96 -3.6% -3.9% -7.4% -8.2% -0.8%Nahlaot, Zichronot 7.93 -5.2% -6.5% -7.8% -8.3% -2.6%Rehavya (AC1) 7.98 -5.7% -7.3% -8.6% -8.9% -3.2%Romema 8.24 -1.8% -2.2% -0.8% -3.1% -2.7%Givat Shaul 7.97 -2.0% -2.2% -1.4% -6.7% -0.5%Har Nof 7.62 -1.5% -1.9% -0.6% -5.1% -1.8%Qiryat Moshe, Bet HaKerem (AC2) 7.67 -2.9% -3.4% -4.7% -6.2% -0.5%Nayot 7.71 -3.1% -3.4% -5.1% -6.6% -0.9%Bayit VaGan 7.86 -3.0% -3.3% -6.0% -7.4% -0.9%Ramat Sharet, Ramat Denya 7.71 -2.4% -2.5% -6.9% -7.3% -0.5%Qiryat HaYovel North 7.78 -3.4% -3.5% -6.7% -7.3% -0.7%Qiryat HaYovel South (NAP3) 7.72 -3.3% -4.7% -7.0% -7.4% -1.2%Qiryat Menahem, Ir Gannim 7.86 -3.8% -3.9% -7.2% -7.7% -0.3%Manahat slopes 7.34 -1.7% -1.8% -4.6% -4.7% -0.5%Gonen (Qatamon) 7.41 -1.1% -1.4% -5.2% -5.4% -0.8%Rassco, Givat Mordekhay 7.44 -1.8% -3.2% -5.4% -5.6% -1.6%German Colony, Gonen 7.28 -1.3% -1.5% -4.1% -4.3% -0.6%Qomemiyyut (Talbiya), YMCA 7.75 -3.0% -3.4% -7.4% -7.7% -0.8%Baqa, Abu Tor, Y. Moshe (AC3) 7.28 -1.1% -1.2% -4.1% -4.2% -0.3%Talpiot, Arnona, Mekor Haim 7.21 -0.5% -0.6% -3.4% -3.6% -0.2%East Talpiot 7.19 -0.9% -1.0% -3.1% -3.3% -0.2%East Talpiot (East) 7.23 -1.2% -1.3% -3.5% -3.7% -0.2%Homat Shmuel (Har Homa) 7.14 -1.3% -1.3% -2.6% -2.8% -0.1%Gilo East 7.55 -2.4% -2.4% -6.4% -6.6% -0.2%Gilo West 7.55 -2.7% -2.8% -6.3% -6.6% -0.2%Talpiot (CD1) 7.14 0.0% 2.3% -2.6% -2.8% -0.2%Givat Shaul (CD2) 7.51 -1.1% 0.8% -1.9% -4.7% -0.4%Malcha CD 7.29 -1.4% -1.4% -4.2% -4.2% -0.3%Romema CD 8.34 -3.5% -6.7% -3.1% -6.5% -1.0%Central Bus Station CD 8.06 -4.2% -4.5% -6.3% -7.0% -1.1%Mahane Yehuda CD 7.79 -4.7% -5.5% -7.1% -7.6% -2.1%

Mean -2.5% -2.8% -4.7% -5.8% -1.0%Median -2.5% -2.8% -4.8% -6.2% -0.8%APP levels

Mean price 7.72 7.53 7.50 7.36 7.27 7.65Median price 7.75 7.51 7.43 7.28 7.20 7.67

Notes: The table reports the percentage changes in Average Prices Paid (APP) charged in all 46 neighbor-hoods. See text for detailed explanations of each scenario. All counterfactuals performed in the third sampleperiod. The last two rows report statistics on the expected prices in levels rather than in percentage changes.

57

F Online appendix: Model, estimation and identifica-

tion: additional details

In this online appendix, we provide some additional technical details regarding the demand

model and its application, including some additional discussion of several aspects of our as-

sumptions.

Deriving equation (2). It is convenient to rewrite the utility function as

Uhjsn = γ−1 ln yj · xjα + δjsn + ζhn(σ) + (1− σ)εhjsn,

where δjsn = νc + νj + νn + hpj · νn − ln psn · xjα − djn · xjβ + κ · hjn is the mean utility

level, common to all origin-j residents who shop at s in destination n. The model is completed

by specifying the utility of a resident of neighborhood j from shopping at the outside option

n = 0, defined as the only member of its nest:

(5) Uhjs0 = γ−1 ln yj · xjα + ζh0(σ) + (1− σ)εhjs0

This definition normalizes, without loss of generality, j-residents’ mean utility from the

outside option at δj0 = 0. The terms vj in the mean utility δjsn associated with “inside options”

allow for heterogeneity in the utility from the outside option across origin neighborhoods. This

is particularly important given that, for residents of neighborhoods in which the price is not

observed, the choice to shop in their home neighborhood is considered part of the outside

option.

The model implies predicted values for choice probabilities and expenditures. Integrating

over the Type I Extreme Value density of the i.i.d. idiosyncratic terms delivers the familiar

nested logit formula for the probability that a resident of neighborhood j shops at store s

located in neighborhood n, conditional on shopping at n,

(6) πjs/n(p; θ) = e(γ−1 ln yj ·xjα+δjsn)/(1−σ)/Djn

where θ = (α, β, κ, σ) are the model’s parameters, and the term Djn is defined by

Djn =Ln∑s=1

exp((γ−1 ln yj·xjα+δjsn)/(1−σ)) for n = 1, ..., 15, and Dj0 = exp(γ−1 ln yjxjα/(1−σ)),

where Ln denotes the number of retailers located in neighborhood n.

The probability that a resident from origin j shops in neighborhood n (the “nest share”) is,

58

(7) πjn(p; θ) = D1−σjn /

N∑m=0

D1−σjm

The probability of shopping at store s located in neighborhood n is given by multiplying

the terms in (6) and (7). Imposing within-neighborhood price symmetry (Assumption 1), we

have psn = pn, and the terms simplify to

Djn = Ln · exp((γ−1 ln yj · xjα + δjn

)/(1− σ))

πjs/n(p; θ) = 1/Ln(8)

πjsn(p; θ) = πjn(p; θ)/Ln

We further obtain that each store in the neighborhood is visited with equal probability so

that demand per neighborhood-j household for the composite good sold at destination n is

(9) qhjn = γ(yj/pn)

Finally, we note that the expected monetary expenditure of household h residing in neighbor-

hood j in destination neighborhood n at time t can be written as ehjnt = πjntqhjntpnt = πjntγyj,

using (9) and taking income to be time-invariant. Because income is assumed identical across

households within the neighborhood, qhjnt and ehjnt do not vary within the neighborhood, and

aggregate expenditures by neighborhood j residents in neighborhood n are,

(10) Ejnt = Hjehjnt = Hjπjntγyj

where Hj is the number of households residing in neighborhood j.46

Motivated by the within-neighborhood store symmetry, we pursue a variant of Berry’s (1994)

inversion strategy: rather than inverting a product (in our case, store) level market share

equation, we invert a nest-level expenditure share equation that equates the nest expenditure

shares predicted by the model to those observed in the data. This enables us to solve for the

mean utility level. Using (7), (10) and the definition of the mean utility δjn, we obtain:47

46We could allow income to vary within neighborhoods by implementing the computationally intensive Ran-dom Coefficient Logit (Berry, Levinsohn and Pakes 1995) instead of the Nested Logit model. We favor thesimplicity of the Nested Logit, particularly in this case since it still allows us to capture the very rich cross-neighborhood variation available in our data.

47Note that the time fixed effect vt is part of the definition of δjnt. Again, the model in Section 3.1 omittedall time indices for expositional clarity.

59

ln

(EjntEj0t

)= ln(πjnt/πj0t) = (1− σ) lnLn + δjnt

= νc + νj + (νn + (1− σ) lnLn) + hpj · νn + νt − ln pnt · xjα− djn · xjβ + κ · hjn

which is equation (2). As shown in the main text, adding Assumption 2 allows us to obtain

the estimation equation (3) which is the one taken to the data.

Identification. The distance effect in the utility function is captured by djn · xjβ, where

xj contains a constant, and shifters such as the origin-j share of car ownership. The coefficient

on the constant term is obtained by relating the variation in expenditures (net of origin, desti-

nation, time and distance effects) in location n to the variation in the distance to n from origin

neighborhoods sharing identical demographics. The other elements of β are identified by relat-

ing this net expenditure variation to the variation in demographics across origin neighborhoods

sharing an identical distance to n.

The price effect is captured by ln pnt · xjα where, similarly, xj contains a constant, and a

shifter of origin-j’s price sensitivity, namely, housing prices. Identification of the constant term

is obtained by relating the net variation in expenditures to the variation in price over time

in the same destination neighborhood. The additional element of α is identified by relating

the net variation in expenditures at destination n to the variation in demographics across

neighborhoods. Note that since we have multiple observations on expenditures in destination

n and from origin j, we could estimate destination and origin fixed effects even with a single

sample period.

Demand elasticities. Demand for the composite good at store s located in neighborhood

n from households residing in neighborhood j is Qjsnt = (Ejsnt/psnt) = Hjπjsnt(γyj/psnt),

where Ejsnt is the total expenditure of origin neighborhood j’s residents at store s located

in neighborhood n and πjsnt is the probability that a resident from origin j shops at the

store. Aggregate demand at the store from all origin neighborhoods is Qsnt =∑J

j=1 Qjsnt. The

retailer’s own price elasticity is therefore

(11) ηsnt,p =psntQsnt

∂Qsnt

∂psnt= −

J∑j=1

Qjsnt

Qsnt

[1 + xjα

(1

1− σ− σ

1− σπjs|nt − πjsnt

)]

where πjs|n was defined in (6). This elasticity measures the percentage change in demand

at store s located in destination n in response to a one percent increase in the composite

good’s price charged at that store. This is a quantity-weighted average of origin-specific price

elasticities.

Imposing the within-neighborhood symmetry mean utility levels (Assumption 1) simplifies

this elasticity term: we obtain πjs|nt = 1/Ln, πjsnt = πjnt/Ln, and Qjsnt/Qsnt = Qjnt/Qnt, where

60

we have denoted the total demand faced by all retailers in neighborhood n Qn, whereas Qjn

is the part of this demand generated by residents of origin j. In other words, the symmetry

assumption implies that the fraction of sales at store s that are made to customers arriving

from neighborhood j is equal to the fraction of total sales by neighborhood n’s retailers to

origin j’s residents. This gives rise to the elasticity formula in (4) as presented in the main

text. Similar calculations deliver the distance semi-elasticity:

ηjnt,d =1

Qjnt

∂Qjnt

∂djn= −xjβ (1− πjnt) ,

measuring the percentage change in demand from residents of neighborhood j at destination

n 6= j in response to a 1 km increase in the distance between these neighborhoods.

Choice probabilities and expenditure shares. In our application πjnt does not neces-

sarily equal the observed expenditure share due to the measurement error and the fact that the

estimated fixed effects (φ) confound the utility fixed effects (ν) with measurement error effects.

As a consequence, even though the parameters α, β, κ are consistently estimated given the as-

sumptions of Section 3.1, the mean utility levels δ are not identified, and hence, neither are the

choice probabilities, absent additional assumptions. Applying the definition Eccjnt =

τjnt

λjntEjnt for

every destination n, and using (10), observed expenditure shares sCCjnt (in words: the share of

expenditures by residents of origin j spent in destination n) can be expressed as:

sCCjnt =Eccjnt∑N

m=0 Eccjmt

=

(τjntλjnt

)πjnt∑N

m=0

(τjmt

λjmt

)πjmt

If, for any fixed origin neighborhood j, the ratio (τjnt/λjnt) is constant across destinations

n, then these ratios cancel out, implying that the observed credit-card expenditure share sCCjnt

is equal to the choice probability πjnt,

(12) sCCjnt =πjnt∑Nm=0 πjmt

= πjnt

This explains the role played by Assumption 3.

The supply side model. We provide here some more detail on the implications of As-

sumption 4 which captures all our assumptions regarding retailers’ behavior. In what follows

we omit the time index t everwhere.

Given rival prices p−sn, the price psn charged by retailer s in destination neighborhood n

maximizes the profit function, Πsn = (psn − cn)Qsn(psn; p−sn), where Qsn =∑J

j=1Qjsn is the

total quantity sold by retailer s in neighborhood n. Rearranging yields the familiar inverse

elasticity formula for the equilibrium margins,

61

(13)psn − cnpsn

= − 1

ηsn,p=

1∑Nj=1

Qjsn

Qsn

[1 + xjα

(1

1−σ −σ

1−σπjs|n − πjsn)]

where the last equality follows from (11).

We follow the literature by assuming the existence of a unique interior Nash equilibrium in

prices.48 We further assume that the unique pricing equilibrium satisfies within-neighborhood

symmetry, a natural assumption given the assumed symmetry of the non-price components of

mean-utility levels. When generating counterfactuals we will compute such equilibria at the

estimated parameter values. It follows that when exploring equilibrium outcomes, we use (8)

to replace πjs|n by 1/Ln, πjsn by πjn/Ln. As explained above in the derivation of the demand

elasticities, this symmetry also allows us to replace (Qjsn/Qsn) by Qjnt/Qnt .

Margins are intuitively affected by within-neighborhood competition, by neighborhood de-

mographics, and by spatial frictions. With respect to within-neighborhood competition, note

that higher values of Ln are associated with lower markups, and the magnitude of this effect

depends on the parameter σ: the derivative of the margin with respect to σ is negative (as long

as Ln > 1). Higher values of σ imply greater substitutability of stores within a neighborhood.

The text offered additional discussion of the intuition underlying the margins formula.

Discussion: some implications of our modeling assumptions. We next provide a

point-by-point discussion of some additional aspects of our assumptions.

1. Complete information. We have implicitly assumed that consumers are perfectly

informed regarding all shopping locations and the prices and amenities offered there. This

stands in contrast to a familiar “search cost” literature in which price differentials are ex-

plained as a consequence of consumers being imperfectly informed about prices (Stigler, 1961).

In Jerusalem, prices in residential neighborhoods are persistently higher than those in the com-

mercial areas. The exact location of the low price stores is common knowledge. This is likely

to be true in many urban settings, and we thus choose to ignore potential information frictions

and emphasize spatial frictions instead.49

2. A single shopping trip. Our model would be misspecified if many consumers split

their grocery shopping among multiple destinations. While such behavior can definitely be

expected, we believe that the time and effort involved with grocery shopping imply that most

consumers perform a single weekly sopping trip, possibly complemented by small “top-up” trips

to make up for a small number of necessary items.

If consumers favor visiting a commercial district where they can split their shopping across

multiple supermarkets, the model would again be misspecified, as it does not allow supermarkets

48Caplin and Nalebuff (1991) demonstrate such uniqueness under stronger conditions than those imposedhere. See also Nocke and Schutz (2015).

49Dubois and Perrone (2015) offer a different view. Other examples of empirical studies of imperfect infor-mation settings include Sorensen (2000), Lach (2002), Brown and Goolsbee (2002), and Chandra and Tapatta(2011).

62

to serve as complements. Most consumers, however, are not likely to split their grocery shopping

across two stores within a single shopping trip. Moreover, greater product variety in shopping

areas is controlled for by the destination fixed effects νn.

Finally, a scenario that would violate Assumption 3 is that households may use credit cards

in their major shopping trip, and cash in small “top-up” trips, performed close to home. In this

case, our measurement error would be correlated with distance, even after controlling for fixed

effects.50 However, as long as the “top-up” trips primarily take place in the home neighborhood,

this issue can be overcome by altering Assumption 3 to condition not only on origin, destination

and time fixed effects, but also on the “shopping at home” dummy variable hjn. This will not

change our estimated coefficients but would change the interpretation of the “shopping at home”

coefficient, which would then confound the utility effect κ with measurement error.

3. Additional unobserved heterogeneity. Our model and estimation follow familiar

strategies in the IO literature based on Berry’s (1994) inversion strategy for the estimation of

demand functions using aggregate data. While we explicitly model measurement error and use it

to construct the econometric error term, the standard approach typically ignores measurement

error and derives the econometric error term by specifying an unobserved random shifter at

the product level. In our context, this would imply adding an unobserved utility shifter vjnt to

equation (3), which would be known to firms and therefore correlated with prices, generating

an identification problem.

The presence of vjnt would imply that residents of certain origin neighborhoods j have a

systematic preference for traveling to certain destination neighborhoods n, over and above the

overall tendency to travel to n (which is controlled for by the vn fixed effect), and for reasons

not related to the distance djn or to the price at the destination pn. We do not expect such

systematic tendencies to be important. One scenario that could generate such tendencies is

that residents of affluent origin neighborhoods may prefer shopping at specific destinations if

those offer unobserved amenities that are particularly appealing to wealthy individuals (e.g.,

better product variety, organic food etc.). We included the term hpj ·νn (origin’s housing prices

interacted with destination fixed effects) to control for such possibilities. This inclusion has

little bearing on the estimated coefficients, reinforcing our prior beliefs that such systematic

effects, to the extent that they are present, are not likely to be quantitatively important.

G Online appendix: Computational details on counter-

factuals

We solve for counterfactual price equilibria, focusing on equilibria that satisfy within-neighborhood

price symmetry. It follows that the pricing equilibrium is characterized by a system of first-order

conditions, containing one “representative” first-order condition per destination neighborhood.

50We are grateful to Pierre Dubois for pointing out this possibility.

63

This is the FOC that characterizes the optimal pricing decision of a representative retailer in

the neighborhood, as defined in (13). It is convenient to organize the FOCs in vector form:

(14) (p− c) • d(p) = p

where • represents element-by-element multiplication and d is a vector defined by

d(p) =

∑J

j=1Qj

1

Q1

[1 + xjα

(1

1−σ −σ

1−σ (1/L1)− πj1/L1

)]∑Jj=1

Qj2

Q2

[1 + xjα

(1

1−σ −σ

1−σ (1/L2)− πj2/L2

)]...∑J

j=1

QjN

QN

[1 + xjα

(1

1−σ −σ

1−σ (1/LN)− πjN/LN)]

The system of equations in (14) is solved by the price equilibrium vector p (assumed to be

unique per discussion above). In each counterfactual experiment, we vary the relevant primitives

and then compute the vector p that solves (14), i.e., the counterfactual price equilibrium. To

perform the counterfactual exercise, one must be able to compute the left hand side of (14),

namely (p− c)•d(p) given any price vector p. Computation of (p− c) is, of course, trivial since

p is given and c is held fixed during the exercise. The critical task is, therefore, the computation

of d(p). Examining the terms inside this vector, we note that xj (observed data) and α (an

estimated parameter) are also held fixed. The terms that need to be calculated are then the

choice probabilities πjn(p), and the quantities Qjn(p)/Qn(p) for each j and n. We now explain

how these are calculated.

We begin by explaining how to calculate πjn(p) for any j, n and a generic value for p. Recall

that the model implies equation (7):

πjn(p; θ) =D1−σjn∑

m∈ND1−σjm

where θ = (α, β, κ, σ) are the model’s parameters, and the term Djn is defined by:

Djn =∑s∈n

e(δjsn+γ−1 ln yjxjα)/(1−σ)

Imposing price symmetry within the neighborhood (which, again, holds by assumption in

the observed equilibrium and in any counterfactual equilibrium), we can write

Djn = e(γ−1 ln yjxjα)/(1−σ) · Ln · e(δjn)/(1−σ)

where, again, Ln denotes the number of symmetric retailers located in neighborhood n, and

the symmetric mean utility is

64

δjn = νc + νj + νn + hpj · νn − ln pn · xjα− djn · xjβ + κ · hjn

The choice probability simplifies to:

(15) πjn(p; θ) =L1−σn eδjn∑

m∈NL1−σm eδjm

To compute these probabilities in the various counterfactuals we need estimates of the mean

utility levels δjn. While the terms ln pn ·xjα, djn ·xjβ and κ ·hjn are known to us given the data,

the estimated parameters and the current guess for p, the terms vc, vj and vn are not known

to us, since the fixed effects actually used in estimation are the terms φj, φn. In other words,

unlike typical applications, our treatment of measurement errors implies that our estimation

strategy does not deliver estimates that allow the direct computation of the mean utility terms

δjn given any price vector.

This, however, is once again resolved given Assumption 3. As shown in Online Appendix F,

this assumption implies that the choice probabilities in the observed equilibrium are equivalent

to the observed credit card expenditure shares. We can use this fact, along with the inversion

principle from Berry (1994), to calculate the mean utility levels δjn in the observed equilibrium.

We then hold these values, denoted δobsjn , fixed and calculate the counterfactual level of δjn, given

any price vector p, by δjn(p) = δobsjn −xjα(ln pn− ln pobsn ). Counterfactuals that change distances

or demographics are handled similarly by appropriately adjusting the observed mean utility

levels.

To compute δobsjn for all j and n, we first recall a result derived in Online Appendix F,

ln

(EjnEj0

)= (1− σ) lnLn + δjn

We further note that

EjnEj0

=Eccjn(λjn/τjn)

Eccj0(λj0/τj0)

=Eccjn

Eccj0

where the first equality holds by definition, and the second equality follows from Assumption

3. We can now obtain an estimate for δobsjn

δobsjn = ln(Eccjn/E

ccj0)− (1− σ) lnLn

where σ = 0.7 is our estimate for the correlation parameter σ. It is, therefore, easy to calculate

δobsjn for all j and n. This enables, as explained above, the calculation of δjn(p) given any price

vector, and the calculation of πjn(p) then follows easily from (15).

It remains to show how to calculate Qjn(p)/Qn(p) for each j and n and any price vector

65

p. Note first that Qjn(p) = Hjπjn(p)qjn = Hjπjn(p)γyj/pn, and that Qn(p) =

N∑j=1

Qjn(p). As a

consequence, we have:

(16) Qjn(p)/Qn(p) =

Hjπjn(p)γyj/pnN∑τ=1

Hτπτn(p)γyτ/pn

=γyjHjπjn(p)

N∑τ=1

γyτHτπτn(p)

We next note that, in the observed equilibrium, the following identity holds: Eccjn = (τjn/λjn)Ejn,

where Eccjn are the observed credit card expenditures. Substituting in the definition of Ejn, we

get that Eccjn = (τjn/λjn)Hjejn = (τjn/λjn)Hjπ

obsjn γyj, implying that:

γyjHj =(λjn/τjn)Ecc

jn

πobsjn

By Assumption 3, the ratio (τjn/λjn) is fixed over all j and n. Substituting into (16), we

then get:

Qjn(p)/Qn(p) =

Mjn · πjn(p)N∑s=1

Msn · πsn(p)

where Mjn = Eccjn/π

obsjn .

Mjn is treated as a constant which is easy to calculate since Eccjn is observed and πobsjn = sccjn.

Since sccjn = Eccjn/

N∑τ=1

Eccjτ , we finally get that Mjn =

N∑τ=1

Eccjτ . That is, this constant is equal

to the total observed expenditures by residents of location j and does not actually vary by n,

that is, Mjn = Mj =N∑τ=1

Eccjτ . The M constants are therefore computed from direct data and

are held fixed during the iterative process that solves the FOCs. The other terms that appear

in Qjn(p)/Qn(p) are choice probabilities πjn(p), and we already explained above how to obtain

those given any p. As a consequence, the final form of d(p) is:

d(p) =

N∑j=1

Mj ·πj1(p)N∑

s=1Ms·πs1(p)

[1 + xjα

(1

1−σ −σ

1−σ (1/L1)− πj1/L1

)]N∑j=1

Mj ·πj2(p)N∑

s=1Ms·πs2(p)

[1 + xjα

(1

1−σ −σ

1−σ (1/L2)− πj2/L2

)]...

N∑j=1

Mj ·πjN (p)N∑

s=1Ms·πsN (p)

[1 + xjα

(1

1−σ −σ

1−σ (1/LN)− πjN/LN)]

66