Retail Prices in a City∗

Alon EizenbergThe Hebrew University and CEPR

Saul LachThe Hebrew University and CEPR

Merav YiftachIsrael Central Bureau of Statistics

December 2019

Abstract

We study grocery price differentials across neighborhoods in a large metropolitan area(the city of Jerusalem, Israel). Using CPI data on prices and neighborhood-level credit carddata on expenditure patterns, we document important variation across neighborhoods intheir access to affordable grocery shopping. Residents of peripheral, non-affl uent neighbor-hoods are charged some of the highest prices in the city, and yet display a low tendency toshop outside their neighborhood. In contrast, residents of affl uent, centrally-located neigh-borhoods often benefit from lower grocery prices charged in their own neighborhood, whilealso displaying a high propensity to shop at the hard-discount grocers located in the city’scommercial districts. We study the role of spatial frictions in shaping these patterns withina structural model where households determine their shopping destination and retailerschoose prices. The estimated model implies strong spatial segmentation in households’demand. In counterfactual analyses, we find that alleviating spatial frictions results inconsiderable benefits to the average resident of the peripheral neighborhoods. At the sametime, it barely affects the equilibrium prices charged across the city, and so it does little tobenefit households with limited mobility (e.g., the elderly).

∗We thank Eyal Meharian and Irit Mishali for their invaluable help with collecting the price data and with theprovision of the geographic (distance) data. We also wish to thank a credit card company for graciously providingthe expenditure data. We are also grateful to Daniel Felsenstein for providing the housing price data, to ElkaGotfryd for mapping zipcodes into statistical subquarters, and to Ruthie Harari-Kremer for her help with maps.We thank Steve Berry, Pierre Dubois, Phil Haile, JF Houde, Gaston Illanes, Volker Nocke, Kathleen Nosal, MarkRysman, Katja Seim, Avi Simhon, Konrad Stahl, Yuya Takashi, Ali Yurukoglu and Christine Zulehner for helpfulcomments, as well as 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). Correspondence: alon.eizenberg@mail.huji.ac.il (Eizenberg), Saul.Lach@mail.huji.ac.il (Lach),meravo@cbs.gov.il (Yiftach). This project was supported by the Israeli Science Foundation (ISF) grant 858/11,by the Wolfson Family Charitable Trust, and by the Maurice Falk Institue for Economic Research in Israel.

1

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-affl uent neigh-

borhoods 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 socioeconomic

factors and geographic location. In this paper we systematically explore these relationships. Us-

ing 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 house-

holds’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-affl uent, peripheral neigh-

borhoods 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 affl uent, centrally-located neighbor-

hoods 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 consis-

tent 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

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 (a Jerusalemlocal newspaper), June 2017.

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

1

groceries. Our demand model quantifies households’price-distance trade-offwhich is substantial:

“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’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 households’

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 expected price paid by a random resident

declines significantly, and this effect is particularly strong for residents of peripheral, non-affl uent

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-

affl uent 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 and counterfactual experiments. Section 5 concludes.

Related literature. Retail price differentials across neighborhoods have attracted consid-erable attention in the economic literature. Such differentials suggest that standard measures

4“The plan: the Talpiot industrial zone to undergo a revolution in the next decade,”Kol Hair, April 2016.

2

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

statistical comparisons of retail prices across affl uent and non-affl uent 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 competi-

tion among supermarkets (for example, Chintagunta, Dubé, and Singh 2003, and Smith 2004).

Our work shares several features with Dubois and Jódar-Rosell (2010) who study supermarket

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 offi cial 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

5MacDonald and Nelson (1991), for example, compared the price of a fixed basket of goods across 322 super-markets in 10 metropolitan areas in the US, revealing that prices in suburban locations were about 4 percentlower than in central city stores where poorer population lived. Chung and Myers (1999) similarly report thatthe price of a weekly home food plan was higher in poorer neighborhoods of the Twin Cities metropolitan area.Recent work, in contrast, reports that prices in richer zip codes (Hayes, 2000) or prices paid by high incomehouseholds (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 Griffi th et al. (2009) for analyses of survey data where recorded pricescorrespond to prices actually paid by households.

3

consumer flows from scanner data, it is not obvious that doing so will provide suffi cient 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.

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 Appendix C and Table C1 for

definitions and a list of neighborhoods). 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. In 2008,

about 65 percent of Jerusalem’s 763,600 residents were Jewish.

Figure 1 displays the neighborhoods, highlighting those covered by our study. 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 neighborhood’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, while Appendix Table C2 shows

neighborhood-specific values.

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.

4

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

than twice as expensive in the central, affl uent neighborhood of Rehavya than in the peripheral,

non-affl uent 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 neighbor-

hoods.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 CD2,

respectively). It also reports statistics on the average distance to all the other neighborhoods, a

rough measure of how peripheral the neighborhood is.9 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-Affl uent 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 Affl uent and Centrally located: Rehavya, Qiryat

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

tively. 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, affl uent and less affl uent 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.

8In online Appendix C we explain that each neighborhood is comprised of several “statistical areas.”The CBSprovided us with the shortest road distance between the centroids of each pair of such areas. We then computethe distance djn between neighborhoods j and n as an average of the distances between each pair of statisticalareas belonging to these neighborhoods. As some neighborhoods are quite large, we define neighborhood j’s “owndistance”djj as the mean distance between the centroids of each pair of statistical areas included in it.

9Appendix Table C3 reports neighborhood-specific values.

5

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 super-

markets, 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 uni-

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

size, packaging, etc.).10 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.11

The list of products, their mean price and coeffi cients of variation are displayed in Tables

D1 and D2 (Online Appendix D). Fruits and vegetables usually exhibit higher price dispersion

than other foodstuff. 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 priceis 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

10Even among fruits and vegetables there are no noticeable quality differences across stores because the CBScollects prices of produce of a specific quality grade.11In our econometric demand model, we address the presence of neighborhoods without sampled prices in an

internally-consistent fashion by including such shopping destinations in the households’outside option.

6

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

pnt =∑i∈Ωnt

(ωi∑

i∈Ωntωi

)pnit. (1)

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

offi cial 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.12 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

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.13

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 (Online Appendix Table D3). In

fact, most of the products appear in all three periods in most neighborhoods (Online Appendix

Table D4). 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 (Online Appendix Table D5).

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 coeffi cients to impute prices in the neighborhoods with missing

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

12While our 27 items are popular products that should be available in all stores, recall that a product is definedby its unique UPC. Some stores may carry a different version of what is essentially the same product (e.g., differingin packaging), generating a missing price observation.13The resulting subsample keeps essentially the same distribution of store formats as the 26 neighborhood

sample (57 percent supermarkets, 21 percent market stalls and 12 percent grocery stores).

7

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 meth-

ods 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 esti-

mated 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

(see Appendix A).14

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

Appendix Table D6 for prices in all neighborhoods). 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 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, Figure 2 shows the distribution of the percentage 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 5-6.5

percent lower in the commercial districts. Variation between residential neighborhoods is also

considerable: panel B shows that prices in our three NAP (non-affl uent peripheral) neighborhoods

are typically ranked above prices in the three AC (affl uent, central) neighborhoods.16 This

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

14In our demand model we partially compensate for the uniform weights by allowing households to derive utilityfrom unobserved aspects of the shopping destinations captured by fixed effects, and by interacting those with theorigin’s housing prices. We therefore allow the variety of additional products (i.e., beyond our 27) to be valueddifferently by households of different income levels.15The composite good’s price increased by 10 percent between November 2007 and November 2008. For

comparison, the CPI inflation for food between December 2007 and December 2008 was 8.3 percent.16AC1, the most affl uent neighborhood is, however, usually more expensive than the NAP neighborhoods.

8

Figure 3 displays the composite good prices in November 2008. It corroborates the observation

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

affl uent neighborhoods. Neighborhoods such as AC2 or AC3 are much more affl uent, 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.

Figure 4 plots composite good prices against housing prices, along with a fitted regression line.17

Prices at AC1, for example, are very high – but are perfectly aligned with that neighborhood’s

affl uence level. Prices at the NAP neighborhoods, in contrast, are considerably higher than what

can be systematically associated with their affl uence level. In our structural model, we will link

these findings to the presence and effects of spatial frictions.18

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 focus on

spatial, rather than informational frictions.19

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.20 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

17Commercial 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 the smallnumber of data points (15 observations) and the lack of other controls preclude us from reaching any conclusionsas to whether “the poor pay more.”18While our framework emphasizes differences in absolute prices charged across neighborhoods with differ-

ent incomes, 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, thecorrelation between the simple composite price index used in the paper and the GEKS-Fisher index is very high(the correlation coeffi cient is 0.89, 0.81 and 0.94 in periods 1-3, respectively).19The rise in the average price in the commercial districts in the third period is entirely due to an unexplained

jump 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.20Credit 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).

9

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 prod-

ucts 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 flowsbetween 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.21 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.22

Table 5 provides statistics regarding the expenditure data (Appendix Table D7 provides

neighborhood-level statistics). The most popular commercial district is CD1 where, on aver-

age, 27 percent of expenditures are incurred. CD1 is the top destination for residents of 16 —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 neighborhood, 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 on the period).

As panel B indicates, residents of the non-affl uent 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 Subsection 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.

21This required a nontrivial mapping between zipcodes and neighborhoods, where zipcodes can map into mul-tiple neighborhoods. We employed a “majority rule”: the zip code was mapped to the neighborhood with whichit has the largest geographical overlap.22We also observe total expenditures of each origin neighborhood at destinations outside the city. Jerusalem does

not have substantial ring of satellite cities providing attractive shopping opportunities. We therefore conjecturethat much of the observed shopping outside the city corresponds to individuals who have a mailing address inJerusalem but do not actually reside in it (e.g., students). We therefore do not use these data in our baselineanalysis. Robustness checks (not reported) in which we added the expenditures incurred outside Jerusalem toour model’s “outside option”(see below) yield remarkably close results to the ones reported in Section 3.2.

10

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 an

“expected price”: the price paid in expectation by a random resident of a neighborhood taking

into account the probability of shopping in each 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 list 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 asumptions (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.

Assumption 1. A household residing in origin neighborhood j = 1, ..., 46 may shop for the

composite good in one of the n = 1, ..., 15 destinations where the composite good prices are

observed, or at the outside option n = 0 corresponding to the remaining 31 neighborhoods.

Assumption 2. Omitting the time index, the (indirect) utility of household h residing in neigh-borhood 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 and νj and νn are origin and destination neighborhood fixed effects, re-

spectively. hpj and yj are, respectively, housing price and average income in origin neighborhood

j; psn is the composite good price at store s located in destination neighborhood n; xj is a vector

of origin neighborhood characteristics; djn is the distance (km) between origin j and destination

n; hjn is a dummy variable equal to 1 if j = n, and zero otherwise. The idiosyncratic term εhjsn

is distributed i.i.d. Type I Extreme Value, and the term ζhn(σ) has the unique distribution that

guarantees that the entire term ζhn(σ) + (1−σ)εhjsn is distributed Type-I Extreme value (Cardell

1997). (γ, a, β, κ, σ) are parameters.

We now discuss the economic content of Assumption 2. The constant νc shifts the utility

from all “inside options”relative to the utility from the outside option, whereas the origin fixed

effects νj capture utility differences across origins with respect to this outside option.23 The23We apply the standard normalization of the outside option’s systematic utility to 0 (Online Appendix F).

11

destination fixed effects νn capture quality differences across destinations which 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 interaction term hpj ·νn allows residents of moreaffl uent neighborhoods to value amenities differently.

The composite good price enters the utility function via the term (γ−1 ln yj − ln psn) · xjα.This functional form follows Björnerstedt and Verboven (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 qhjsn = γyj/psn, so that expenditure on the composite

good is a constant fraction γ of the (representative) household’s income.24 The fraction γ drops

out of the estimation equation and therefore it could vary across origin neighborhoods.25 The

parameter vectors α (β) allow price (distance) sensitivities to vary with origin neighborhood

characteristics.

The “shopping at home”parameter κ 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 introduces nonlinearity in the household’s travel costs: it 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.

Finally, the idiosyncratic term ζhn(σ) + (1 − σ)εhjsn follows the assumptions for the nested

logit model (Berry 1994). The nests are destination neighborhoods allowing stores within a

neighborhood to be closer substitutes than stores located in different neighborhoods. 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. The parameter σ takes values in the

interval [0, 1) and captures the correlation in tastes over stores within a neighborhood. It governs

the intensity of within-neighborhood competition: larger values imply stronger substitutability.

Assumption 3. Denote by δjsn = νc + νj + νn + hpj · νn − ln psn · xjα − djn · xjβ + κ · hjn themean utility level, common to all origin j residents who shop at s in destination n. Stores within

a neighborhood offer identical mean utility levels across households, δjsn = δjn for every j, n.

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

tion 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 4.24Defining f(yj , psn) = (γ−1 ln yj − ln psn), Roy’s identity implies that qhjsn = −∂f/∂psn/∂f/∂yj .25More 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.

12

Assumption 3 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 supermarket nearest to them. It also allows for any

other type of horizontal differentiation among supermarkets 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) and, consequently, as shown in

Section 2.2, most of the price variation is across, rather than within, neighborhoods.26

In Online Appendix F we show that Assumptions 1-3 deliver the following linear equation:

ln

(EjntEj0t

)= νc + νj + (νn + (1− σ) lnLn) + hpj · νn + νt − ln pnt · xjα− djn · xjβ + κ · hjn, (2)

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

tination 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 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 = τ jntEjnt, with 0 ≤ τ jnt ≤ 1. These

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

Eccjnt via E

ccjnt = (τ jnt/λjnt)Ejnt.

Thus, adding wjnt = ln(τ jntλjnt

λj0tτ j0t

)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 mismeasurement

26The 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 effi cient coverage of the shopping probabilities characterizing each origin-destination pair.

13

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 affl uent 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 4. Conditional on origin, destination and time fixed effects, wjnt is uncorrelatedwith prices and distances.

This assumption implies that the proportionality factors may depend on fixed neighborhood

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 4 allows us to rewrite (2) as

ln(Eccjnt/E

ccj0t) = φc + φj + φn + φt + hpj · vn − ln pnt · xjα− djn · xjβ + κ · hjn + ujnt (3)

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.27

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 Assumptions 1-4.

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

27While 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 observationsfrom the sample, reducing its size from 46 × 15 × 3 = 2070 observations to 1819. The results are qualitativelyrobust to substituting a very small number for Eccjnt (Appendix A). This is not a formally valid correction butone often used in practice (see Gandhi, Lu and Shi 2013 for a partial identification approach).

14

is to calibrate σ so that it generates reasonable markups given the identified parameters.28

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.29 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.30 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

of socioeconomic interactions. Across all specifications, the coeffi cients have the expected signs.

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

so is the coeffi cient on the “shopping at home,”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 are strongly 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),

and yields higher estimation precision. This clarifies the importance of controlling for 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 char-

acteristics 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

28Such an approach has some precedence in the literature. For example, Björnerstedt 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 our calibration as it eliminatesthe need to rely on our pricing model in generating the demand estimates.29Note that these are markups above marginal cost. They are, therefore, higher than markups over average

costs, the latter often approximated using information from retailers’financial reports.30We also regressed the estimated fixed effects φn on lnLn to estimate 1 − σ. This yields an (imprecisely

estimated 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.

15

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, allow-

ing for different valuations of unobserved amenities across socioeconomic classes. The estimated

price coeffi cient is mildly reduced (from 5.1 to 4.7). Distance coeffi cients are also only minimally

affected except for the interaction with the percentage of senior citizens. We adopt column (6) as

our baseline specification. Overall, estimates in columns (7) —(10) are very close to the baseline

specification in column (6).31

Elasticities. The economic implications of these estimates are captured in price and distanceelasticities. The own-price elasticity faced by store s in neighborhood n is (Appendix F):

ηsn,p =psnQsn

∂Qsn

∂psn= −

J∑j=1

Qjn

Qn

[1 + xjα

(1

1− σ −σ

(1− σ)Ln− πjnLn

)](4)

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

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 2008.32

31Interacting price with family size in additional, unreported specifications yields an insignificant effect anddoes not alter the other coeffi cients.32The number of supermarkets is shown in the last column of Appendix Table C2. This value includes all

supermarkets, 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 values affectsthe margins for retailers in this specific neighborhood, but does not affect the qualitative findings of the paper.

16

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 5. For each origin j, the ratio τ jnλjn

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 mea-

surement error altogether. We stress that Assumption 5 is not required for the consistent estima-

tion 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 Appendix B, it makes no difference in terms of our

qualitative conclusions.33 Appendix A presents the robustness of the implied 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 consideration

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 for33Further, increasing σ to 0.9 makes demand even more elastic, which is intuitive, but once again does not

affect our qualitative conclusions. Details are available from the authors upon request.

17

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: expected prices

In Section 2.2 we showed that prices in non-affl uent, peripheral neighborhoods were often higher

than prices in more affl uent but centrally located neighborhoods. This, however, does not nec-

essarily 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 5, 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 an expected price for residents of neighborhood j, interpreted as the cost of

grocery shopping incurred by a random resident of origin neighborhood j: pEj ≡∑N

m=0 πjmpm.34

Figure 5 plots the expected price against housing prices in each of the 46 neighborhoods in

November 2008 (along with a linear predicted line). In 8 out of the 11 residential neighborhoods

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

savings afforded to households by shopping outside their home neighborhood (when pEj is higher

than pj, the difference is small).

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

have been the price actually paid if households were to determine their shopping destination based

on price only (ignoring equilibrium effects). The expected price 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 expected prices at the peripheral, non-affl uent neighborhoods NAP1-NAP3

are higher than those faced by residents of more affl uent neighborhoods that are located closer to

the commercial areas, AC2-AC3. Moreover, a strong, positive relationship is depicted in Figure

6 between the expected price pEj 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.34This 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-affl uent neighborhood that launched the consumer boycott in 2014.

18

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.

4 Retail supply: pricing decisions

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

Assumption 6. (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 6, along with the symmetric differentiation assumption from the de-

mand model (Assumption 3) naturally allow us to focus on the (within-neighborhood) symmetric

price equilibria assumed in part (iv).

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

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

tuitively 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 (Appendix B).

19

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-affl uent, 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.35

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)

equilibrium prices (ii) the probabilities with which residents of each origin neighborhood shop at

each destination, and (iii) the prices paid by residents of each neighborhood in expectation.36

We use the baseline estimates from column 6 in Table 6 and σ = 0.7. 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 of the computation of the counterfactual equilibria.

Table 9 summarizes the impact on the prices charged by retailers across the city (Appendix

Table E1 offers complete neighborhood-specific results). The first column corresponds to the

observed 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 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 affl uent, peripheral

35We do not consider multi-store pricing by chains that operate supermarkets in both residential and commercialdistricts. Because these chains’pricing in the commercial districts is strongly constrained by the presence of harddiscounters that operate only there, we expect this issue to have little impact on our findings.36The model allows us to compute the impact on welfare but as our exercises directly affect utility parameters

we find this less appealing.

20

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.37

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.38

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

at the entry points to the commercial district with a convenient shuttle service.39 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 the

NAP1-NAP3).

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

petitors in each residential neighborhood n by 1. The median 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 prices respond very mildly, households’shopping behavior changes markedly:

37This is similar to the observed “generic drug paradox”that occurs when many, but not all, consumers switchto newly available generic drugs but prices among the incumbent (brand) drugs do not decline and even rise(Griliches and Cockburn, 1994).38Given that νn is unidentified due to measurement error, we use the standard deviation of φn, the fixed effect

that 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.39Interestingly, 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).

21

many households switch to shopping in the affordable commercial districts. This is captured by

the expected price, the price faced in expectation by a random resident of each neighborhood.

Table 10 summarizes the counterfactual changes to the expected prices faced by residents of the

eleven residential neighborhoods where prices are observed.40

The first column of Table 10 reports the expected price faced by residential neighborhoods

in the observed equilibrium (corresponding to the data points in Figure 6). The percentage

reduction in expected prices faced by neighborhood residents following 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) expected price falls by 1.8 —5.6 percent,

versus the 0.1 —0.8 percent fall in median prices shown in Table 9.

Expected prices, therefore, respond much more strongly than equilibrium prices 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 expected price 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 expected prices in Table 10 take also into account the changes in shopping patterns

induced by the changes in parameters. This is evident in Figure 7 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 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, expected prices decline considerably.

Viewed through the lens of its effect on expected prices, 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 expected price paid by residents of NAP3

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

10), whereas the price charged by the retailers at NAP3 dropped by 0.6 percent only (Table 9).

Expected prices at NAP1 and NAP2 drop by 2.2 and 6.6 percent, respectively, whereas prices

charged by the retailers in both of these neighborhoods only drop by 0.1 percent.

Discussion. Evaluating the impact of a reduction in spatial frictions by considering only the40For completeness, Appendix Table E2 shows the impact on expected prices for all 46 neighborhoods, delivering

the same qualitative conclusions. We favor presenting here results for the 11 residential neighborhoods whereprices are observed to facilitate comparison with the impact on prices displayed in Table 9.

22

effect on 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 expected prices, 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 substantial

benefits to the average resident of the peripheral, non-affl uent neighborhoods.41 However, because

prices charged at the NAP neighborhoods barely change, or even increase, a reduction 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 households 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 (e.g., food stamps). 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-affl uent neighborhoods often charge higher

prices than retailers located in more centrally located, affl uent neighborhoods.

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 peripheral

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

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

ever, 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

41A caveat to this statement is that the attractiveness of a location vn is probably endogenous and might changeif it experiences a substantial increase in shopping activity. Addressing this would require a model of retailers’choice of amenities which is beyond the scope of this paper.

23

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 (e.g., food stamps) 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 Disper-

sion,”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] Björnerstedt, J., and F. Verboven (forthcoming in 2016): “Does Merger Simulation Work?

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

nomics.

[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. Dubé, and V. Singh (2003): “Balancing Profitability and Customer

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

24

[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. Jódar-Rosell (2010): “Price and Brand Competition between Differenti-

ated 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 fro

25

[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

[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 Economics

10, 141-156

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

rium,”Review of Economic Studies, 71(1), 235—263

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

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

26

7 Tables and Figures

Table 1: Distribution of demographics across 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. Drivingto work = percentage of individuals above 15 years of age who used a private caror a commercial vehicle (as a driver) as their main means of getting to work inthe determinant week. Car ownership = percentage of households using at leastone car. Senior citizens = percentage above age 65.

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

Variable N mean sd min median max

City center 46 4.3 2.3 0.6 4.3 9.2Commercial District 1 (CD1) 46 5.7 2.9 0.0 5.4 13.2Commercail District 2 (CD2) 46 6.0 2.5 0.0 5.8 12.0All destinations (mean) 46 6.1 1.6 4.2 5.8 10.8

Notes: Distribution of distances between neighborhoods and 1) the citycenter, 2) the two prominent commercial centers, and 3) all other neighbor-hoods. Statistics over the 46 neighborhoods are displayed. Source: CBS

27

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-Affl uent Peripheral, and Affl uent Central (seetext). The number of supermarkets (most-right column) includes all supermarkets in the neigh-borhood, not just those where prices were sampled.

28

Table 4: Price of composite good across neighbothoods and time

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

Prices

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 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

Notes: Panel A displays composite good price statistics over the 15 neighborhoods where the price could becomputed using at least 21 observed products (see text). Panel B presents values for particular neighborhoodswhere CD, NAP and AC correspond to Commercial Districts, Non-Affl uent Peripheral, and Affl uent Central (seetext). The last column in panel B presents the neighborhood’s mean price rank over the three sample periods.

29

Table 5: Credit card expenditure flows

A. Statistics over the 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 the frac-tion of expenditures spent at the own neighborhood and at the CD1 and CD2commercial districts. NAP and AC correspond to Non-Affl uent Peripheral, andAffl uent Central neighborhoods (see text).

30

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 (price) X 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 X seniors 0.002 0.004 0.004 0.004 0.004 0.006(.004) (.007) (.007) (.006) (.005) (.007)

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

Distance X 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 X housing prices 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 price and distance variables were entered with a negative sign in the regression so that the estimates in the tableare estimates of α and β. Standard errors in parentheses are 2-way clustered at the origin and destination levels.

31

Table 7: Distribution of estimated elasticities (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. Price elasticities are computedfor each destination. Distance semi-elasticities computed for each origin-destination pair.

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 in each destination neighborhoodin which prices are observed in November 2008. Costs and mar-gins are reported under two alternative values for the correlationparameter σ. CD, NAP and AC correspond to Commercial Dis-tricts, Non-Affl uent Peripheral, and Affl uent Central (see text).

32

Table 9: Percentage change in prices under counterfactual scenarios

Retail location Observed price Reduced travel disutility Improved amenities Additional entryReduced distance disutil. + reduced 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 in prices charged at locations where prices are observed (11residential neighborhoods and 4 commercial districts) under the various counterfactual analyses, computed in the third timeperiod (November 2008). CD, NAP and AC correspond to Commercial Districts, Non-Affl uent Peripheral, and Affl uentCentral (see text).

Table 10: Percentage change in expected prices under counterfactual scenarios

Retail location Expected price Reduced travel disutility Improved amenities Additional entry(observed equil.) Reduced distance disutil. + reduced 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 expected prices (i.e., those that are paid by a random resident in expecta-tion over the shopping destination probabilities) charged at the 11 residential neighborhoods where prices are observed. Seetext for detailed explanations. All analyses performed for the third time period (November 2008). The first column shows ex-pected prices corresponding to the observed equilibrium, whereas additional columns show the counterfactual expected pricesgiven alternative parameter values as described in the text. CD, NAP and AC correspond to Commercial Districts, Non-Affl uent Peripheral, and Affl uent Central (see text).

33

NAP1

NAP2

AC1

AC2

NAP3AC3

CD2

CD1

Kilometers

0 5Not included (16)Included (46)

Figure 1: Neighborhoods included in the study

34

0.1

.2.3

.4Fr

actio

n

0 5 10 15 20Gains from travelling(%)

Figure 2: Histogram of the percentage savings from shopping at the cheapest neighborhood; allthree periods

35

NAP1

NAP2

AC1

AC2

NAP3AC3

CD2

CD1

Kilometers

0 5

Price (NIS)8.00 - 8.697.50 - 8.006.88 - 7.50No data

Figure 3: Composite good prices across the city, November 2008

36

NAP1

NAP2NAP3

AC3

AC2

AC1

CD1

CD2

77.

58

8.5

9

10 15 20Housing prices ('000s NIS)

Residential neighborhood Commercial districtFitted values

Figure 4: Composite good prices plotted against housing prices, November 2008

37

CD1

NAP1 NAP2

NAP3

CD2

AC3

AC2

AC1

77.

58

8.5

10 15 20Housing prices ('000s NIS)

Expected composite good price Fitted values

Figure 5: Observed prices and expected prices plotted against housing prices, November 2008

38

CD1

AC3

AC1

NAP3AC2

CD2

NAP2 NAP1

77.

58

8.5

0 5000 10000 15000Distance to CD1 (in meters)

Expected composite good price Fitted values

Figure 6: Expected prices plotted against distance to CD1, November 2008

39

NAP1

NAP2

AC1

AC2

NAP3AC3

CD2

CD1

Kilometers

0 5

Ratio of probabilities4 - 5.22 - 41.2 - 2No data

Figure 7: Ratio of counterfactual (improved amenities at CD1) to observed probability of shop-ping at CD1, November 2008

40

A Robustness of demand estimates to the computationof 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 coeffi cient and the coeffi cient 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

42The socio economic variables used to predict prices are a subset of the following: number of family households,median age, percentage of married people aged 15 and over, average number of persons per household, percentageof households with 7+ persons in the household, percentage of households with 5+ children up to age 17 in thehousehold, dependency ratio, percentage of those aged 15 and over in the annual civilian labor force, percentage ofthose aged 15 and over who did not work in 2008, percentage of Jews born abroad who immigrated in 1990-2001,percentage of households residing in self-owned dwellings, percentage of Jews whose origin is Israel, percentage ofJews whose continents of origin are America and Oceania, percentage of Jews whose 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 with9-12 years of schooling, percentage of those aged 15 and over with 13-15 years of schooling, percentage of thoseaged 15 and over with 16 or more years of schooling. In addition, we added an indicator for a commercial districtand period dummies. The R2′s of these 27 regressions are quite high, ranging from 0.45 to 0.93 with a medianvalue of 0.70.43In 16 observations with missing prices where the imputed price was negative it was substituted for by the

minimum 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.

41

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 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 coeffi cients (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 coeffi cient between the original composite

44In a few locations, the basket is composed of nine or ten fruits and vegetables.

42

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.

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

43

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 in No-vember 2008. Distance semi-elasticities are computed for each origin-destination pair (e.g., 46x15=690).

44

Table A3: Dermand 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)

45

B Counterfactual analyses for σ = 0.8

Table B1: Percentage change in prices under counterfactual scenarios, sigma=0.8

Retail location Observed price Reduced travel disutility Improved amenities Additional entryReduced distance disutil. + reduced 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 than σ = 0.7. Seenotes to table 9 for additional details.

Table B2: Percentage change in expected prices under counterfactual scenarios, sigma=0.8

Retail location Observed price Reduced travel disutility Improved amenities Additional entry(expected) Reduced distance disutil. + reduced 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 rather than σ = 0.7. Seenotes to table 10 for additional details.

46

C Online appendix: Neighborhoods, subquarters and de-mographics

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 excep-

tions: 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 sub-

quarters 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 de-

mographics 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.

47

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.

48

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. Driving to work = percentage ofthose aged 15 and over who used a private car or a commercial vehicle (as a driver) as their main means ofgetting to work in the determinant week. Car ownership = percentage of households using at least one car.Senior citizens = percentage of those aged 65+ . Source: CBS. The number of supermarkets includes all su-permarkets in the neighborhood, not just those where prices were sampled.

49

Table C3: Distances (in km)

Neighborhood Distance to:

All neighborhoods City CD 1 CD 2(mean) center

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 between neighborhoods and 1) the city center, 2) the two prominent com-mercial centers, and 3) all other neighborhoods. Statistics over the 46 neighborhoods aredisplayed. Source: CBS.

50

D Online appendix: Products, prices and expenditures

Table D1: Definition of products

1 Waffl es simple packed waffl es, 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

51

TableD2:Product-specificpricedistributions(NIS)

Product

Mean

Coefficient

#stores

Product

Mean

Coefficient

#stores

Product

Mean

Coefficient

#stores

price

ofVariation

price

ofVariation

price

ofVariation

Waffl

esTurkishcoffee

Cauliflower

Sep-07

10.4

0.14

24Sep-07

5.8

0.09

23Sep-07

7.3

0.32

25Nov-07

10.2

0.18

22Nov-07

5.7

0.11

23Nov-07

5.9

0.19

22Nov-08

11.1

0.24

20Nov-08

70.07

23Nov-08

6.6

0.24

23

Mayonnaise

Cocoapowder

Potatoes

Sep-07

7.6

0.12

22Sep-07

10.3

0.12

23Sep-07

40.23

37Nov-07

90.21

21Nov-07

10.5

0.12

23Nov-07

4.2

0.26

37Nov-08

9.6

0.14

16Nov-08

10.7

0.11

22Nov-08

4.8

0.25

35

Cottagecheese

Green

peas(can)

Tom

atoes

Sep-07

5.3

0.04

23Sep-07

5.2

0.10

16Sep-07

6.1

0.33

37Nov-07

5.8

0.03

25Nov-07

5.2

0.10

16Nov-07

50.34

37Nov-08

60.05

22Nov-08

5.9

0.12

14Nov-08

6.9

0.33

35

Sugar

Hummus(salad)

Apples

Sep-07

3.6

0.22

24Sep-07

90.11

17Sep-07

90.20

36Nov-07

3.6

0.22

23Nov-07

9.2

0.05

18Nov-07

9.1

0.12

34Nov-08

3.4

0.26

24Nov-08

10.6

0.10

14Nov-08

9.6

0.18

33

Chocolatebar

Cucumbers

Bananas

Sep-07

4.4

0.11

23Sep-07

4.6

0.28

37Sep-07

6.3

0.13

35Nov-07

4.5

0.11

23Nov-07

5.8

0.17

37Nov-07

5.6

0.30

35Nov-08

5.1

0.12

23Nov-08

4.8

0.29

35Nov-08

7.8

0.23

33

Mineralwater

Onion

Lemons

Sep-07

12.8

0.11

21Sep-07

2.8

0.32

37Sep-07

11.7

0.22

38Nov-07

12.7

0.15

20Nov-07

3.2

0.34

36Nov-07

8.1

0.25

36Nov-08

12.3

0.28

20Nov-08

3.7

0.35

35Nov-08

10.4

0.37

35

Cocacola

Carrots

Fabrics.

Sep-07

5.5

0.18

25Sep-07

4.9

0.18

37Sep-07

20.8

0.08

21Nov-07

5.5

0.18

25Nov-07

5.1

0.18

36Nov-07

19.9

0.16

25Nov-08

5.9

0.17

24Nov-08

5.6

0.38

32Nov-08

22.1

0.07

22

Ketchup

Eggplants

Dishwasherd.

Sep-07

11.1

0.14

24Sep-07

40.40

38Sep-07

10.8

0.12

16Nov-07

10.9

0.14

24Nov-07

3.7

0.41

35Nov-07

11.9

0.10

19Nov-08

110.15

23Nov-08

4.7

0.34

33Nov-08

11.1

0.20

23

Tea

Cabbage(white)

Shavingc/g

Sep-07

15.8

0.15

22Sep-07

4.7

0.51

33Sep-07

22.1

0.20

22Nov-07

16.2

0.15

23Nov-07

3.7

0.57

32Nov-07

23.2

0.22

16Nov-08

17.1

0.15

20Nov-08

5.1

0.61

31Nov-08

23.5

0.16

18

52

Table D3: Number of sampled stores and of observed products

Number of sampled stores Number of 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.

53

Table D4: Product composition of composite good

Neighborhood

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15ProductWaffl es 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.

54

Table D5: Distribution of products across neighborhoods

Sep-07 Nov-07 Nov-08

Waffl es 15 13 13

Low 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.

55

Table D6: Composite good prices across neighbothoods and time

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.

56

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.

57

E Online appendix: Observed and counterfactual expectedprices 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 presentes changes in expected prices in all 46

neighborhoods.

58

TableE1:Percentagechangeinpricesundercounterfactualscenarios

Neighborhood

Observedprice

Reducedtraveldisutility

Improvedamenities

Additionalentry

Reduceddist.Disutil.

+reducedκ

CD1

CD1-CD2

NeveYaaqov(NAP1)

8.01

3.3%

4.7%

-0.1%

-0.3%

-2.8%

PisgatZeevNorth

7.36

0.3%

0.7%

-0.9%

-1.2%

-3.4%

RamotAllonnorth

7.61

0.2%

0.3%

-0.4%

-0.5%

-3.3%

GivatShapira(NAP2)

8.14

0.4%

0.8%

0.0%

-0.1%

-1.3%

Rehavya(AC1)

8.52

-8.2%

-12.0%

-3.6%

-1.1%

-6.8%

Romema

8.17

-1.3%

-2.7%

1.2%

1.9%

-4.4%

HarNof

7.62

-0.7%

-1.7%

0.0%

-1.3%

-4.3%

QiryatMoshe,B.HaKerem

(AC2)

7.85

-1.3%

-3.7%

0.2%

-0.8%

-1.9%

QiryatHaYovelSouth(NAP3)

8.19

-0.5%

-0.5%

-0.6%

-0.9%

-3.5%

Rassco,GivatMordekhay

7.87

-1.7%

-3.4%

-0.7%

-0.9%

-4.6%

Baqa,AbuTor,Y.Moshe(AC3)

7.76

-0.2%

-0.3%

-0.1%

-0.2%

-3.0%

Talpiot(CD1)

6.89

-0.3%

-0.2%

0.5%

0.2%

0.0%

GivatShaul(CD2)

7.07

-1.1%

-1.0%

0.3%

0.3%

0.0%

RomemaCD

8.69

-1.0%

-1.0%

0.4%

0.3%

0.1%

MahaneYehudaCD

7.20

-1.5%

-1.4%

0.1%

-0.2%

0.1%

Mean(residential)

-0.9%

-1.6%

-0.5%

-0.5%

-3.6%

Median(residential)

-0.5%

-0.5%

-0.1%

-0.8%

-3.4%

Notes:Thetablereportsthepercentagechangesinpriceschargedinthe15neighborhoodswherethecompositegoodpricecouldbe

computedusingatleast21goods.Thecounterfactuals,performedinthethirdsampleperiod,aredescribedinthetextindetail.

59

Table E2: Percentage change in expected prices under counterfactual scenarios

Retail location Observed price Reduced travel disutility Improved amenities Additional entry(expected) Reduced distance disutil. + reduced 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%Expected price 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 expected prices charged in all 46 neighborhoods. See text for detailedexplanations of each scenario. All counterfactuals performed in the third sample period. The last two rows report statisticson the expected prices in levels rather than in percentage changes.

60

F Online appendix: Model, estimation and identification: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 assumptions.

Deriving equation (2) given assumptions 1-3. It is convenient to rewrite the utilityfunction in Assumption 2 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:

Uhjs0 = γ−1 ln yj · xjα + ζh0(σ) + (1− σ)εhjs0 (5)

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 overthe 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,

πjs/n(p; θ) = e(γ−1 ln yj ·xjα+δjsn)/(1−σ)/Djn (6)

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,

πjn(p; θ) = D1−σjn /

N∑m=0

D1−σjm (7)

61

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 3), 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

qhjn = γ(yj/pn) (9)

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,

Ejnt = Hjehjnt = Hjπjntγyj (10)

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 equa-

tion, 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

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 4 allows us to obtain the

estimation equation (3) which is the one taken to the data.

46We could allow income to vary within neighborhoods by implementing the computationally intensive RandomCoeffi cient Logit (Berry, Levinsohn and Pakes 1995) instead of the Nested Logit model. We favor the simplicityof the Nested Logit, particularly in this case since it still allows us to capture the very rich cross-neighborhoodvariation 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 omited all

time indices for expositional clarity.

62

Identification. The distance effect in the utility function is captured by djn · xjβ, where xjcontains a constant, and shifters such as the origin-j share of car ownership. The coeffi cient on the

constant term is obtained by relating the variation in expenditures (net of origin, destination, 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 relating 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 ashifter 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 neighborhoodn 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 neighbor-

hood 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=1Qjsnt. The retailer’s own price

elasticity is therefore

ηsnt,p =psntQsnt

∂Qsnt

∂psnt= −

J∑j=1

Qjsnt

Qsnt

[1 + xjα

(1

1− σ −σ

1− σπjs|nt − πjsnt)]

(11)

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 3) simplifies

this elasticity term: we obtain πjs|nt = 1/Ln, πjsnt = πjnt/Ln, and Qjsnt/Qsnt = Qjnt/Qnt, where

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

asssumption 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:

63

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 necessarilyequal the observed expenditure share due to the measurement error and the fact that the esti-

mated 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 Assumptions

1-4, the mean utility levels δ are not identified, and hence, neither are the choice probabilities,

absent additional assumptions. Applying the definition Eccjnt =

τ jntλjnt

Ejnt 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

=

(τ jntλjnt

)πjnt∑N

m=0

(τ jmtλjmt

)πjmt

If, for any fixed origin neighborhood j, the ratio (τ jnt/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,

sCCjnt =πjnt∑Nm=0 πjmt

= πjnt (12)

This explains the role played by Assumption 5.

The supply side model. We provide here some more detail on the implications of Assump-tion 6 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,

psn − cnpsn

= − 1

ηsn,p=

1∑Nj=1

QjsnQsn

[1 + xjα

(1

1−σ −σ

1−σπjs|n − πjsn)] (13)

where the last equality follows from (11).

64

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 demo-

graphics, 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 de-

pends 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 in-formed 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 explained 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 commercial 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 theirgrocery 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

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

48Caplin and Nalebuff (1991) demonstrate such uniqueness under stronger conditions than those imposed here.See also Nocke and Schutz (2015).49Dubois and Perrone (2015) offer a different view. Other examples of empirical studies of imperfect information

settings include Sorensen (2000), Lach (2002), Brown and Goolsbee (2002), and Chandra and Tapatta (2011).

65

areas is controlled for by the destination fixed effects νn.

Finally, a scenario that would violate Assumption 4 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 4 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 coeffi cients but would change the interpretation of the “shopping at home”

coeffi cient, 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 affl uent 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 pricesinteracted with destination fixed effects) to control for such possibilities. This inclusion has little

bearing on the estimated coeffi cients, 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

50We are grateful to Pierre Dubois for pointing out this possibility.

66

conditions, containing one “representative”first-order condition per destination neighborhood.

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:

(p− c) • d(p) = p (14)

where • represents element-by-element multiplication and d is a vector defined by

d(p) =

∑J

j=1Qj1Q1

[1 + xjα

(1

1−σ −σ

1−σ (1/L1)− πj1/L1

)]∑Jj=1

Qj2Q2

[1 + xjα

(1

1−σ −σ

1−σ (1/L2)− πj2/L2

)]...∑J

j=1

QjNQN

[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 sincep 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∈N

D1−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)

67

where, again, Ln denotes the number of symmetric retailers located in neighborhood n, and the

symmetric mean utility is

δjn = νc + νj + νn + hpj · νn − ln pn · xjα− djn · xjβ + κ · hjn

The choice probability simplifies to:

πjn(p; θ) =L1−σn eδjn∑

m∈N

L1−σm eδjm

(15)

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 5. 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 thatEjnEj0

=Eccjn(λjn/τ jn)

Eccj0(λj0/τ j0)

=Eccjn

Eccj0

where the first equality holds by definition, and the second equality follows from Assumption 5.

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).

68

It remains to show how to calculate Qjn(p)/Qn(p) for each j and n and any price vector 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:

Qjn(p)/Qn(p) =

Hjπjn(p)γyj/pnN∑τ=1

Hτπτn(p)γyτ/pn

=γyjHjπjn(p)

N∑τ=1

γyτHτπτn(p)

(16)

We next note that, in the observed equilibrium, the following identity holds: Eccjn = (τ jn/λjn)Ejn,

where Eccjn are the observed credit card expenditures. Substituting in the definition of Ejn, we

get that Eccjn = (τ jn/λjn)Hjejn = (τ jn/λjn)Hjπ

obsjn γyj, implying that:

γyjHj =(λjn/τ jn)Ecc

jn

πobsjn

By Assumption 5, the ratio (τ jn/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 π

obsjn = 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:

69

d(p) =

N∑j=1

Mj ·πj1(p)N∑s=1

Ms·πs1(p)

[1 + xjα

(1

1−σ −σ

1−σ (1/L1)− πj1/L1

)]

N∑j=1

Mj ·πj2(p)N∑s=1

Ms·πs2(p)

[1 + xjα

(1

1−σ −σ

1−σ (1/L2)− πj2/L2

)]

...

N∑j=1

Mj ·πjN (p)N∑s=1

Ms·πsN (p)

[1 + xjα

(1

1−σ −σ

1−σ (1/LN)− πjN/LN)]

70