Who’s Getting Globalized? The Size and Nature of
Intranational Trade Costs∗
David Atkin† and Dave Donaldson‡
July 2012
Abstract
This paper uses a newly collected dataset on the prices of narrowly defined goodsacross many dispersed locations within multiple developing countries to address thequestion, How large are the costs that separate households in developing countries from theglobal economy? Guided by a flexible model of oligopolistic intermediation with vari-able mark-ups, our analysis proceeds in four steps. First, we measure total intrana-tional trade costs (ie marginal costs of trading plus mark-ups on trading) using pricegaps over space within countries—but we do so only among pairs of locations thatare actually trading a good by drawing on unique data on the location of productionof each good. Second, we estimate, separately by location and commodity, the pass-through rate between the price at the location of production and the prices paid by in-land consumers of the good. Our estimates imply that incomplete pass-through—andtherefore, intermediaries’ market power—is a commonplace, and that pass-through isespecially low in remote locations. Third, we argue that our estimates of total tradecosts (Step 1) and pass-through rates (Step 2) are sufficient to infer the primitive effectof distance on the marginal costs of trading; after correcting for the fact that mark-upsvary systematically across space we find that marginal costs are affected by distancemore strongly than typically estimated. Finally, we show that, in our model, the esti-mated pass-through rate (Step 2) is a sufficient statistic to identify the shares of socialsurplus (ie the gains from trade) accruing to inland consumers, oligopolistic inter-mediaries, and deadweight loss; applying this result we find that intermediaries inremote locations capture a considerable share of the surplus created by intranationaltrade.
∗We thank Rohit Naimpally for excellent research assistance, and Alvaro González, Leonardo Iacovone,Clement Imbert, Horacio Larreguy, Philip Osafo-Kwaako, John Papp, and the World Bank Making MarketsWork for the Poor Initiative for assistance in obtaining segments of the data. We have benefited greatlyfrom conversations with Pol Antras, Arnaud Costinot, Michal Fabinger, Marc Melitz, Glen Weyl and manyothers, as well as from comments made by seminar participants at Bergen, Berkeley, Harvard, the IGC TradeGroup, the IGC Infrastructure and Urbanization Conference in London, the LSE, MIT, Stanford, UC SantaCruz. Finally, we thank the International Growth Centre in London for their generous financial support.†Yale University and NBER. E-mail: [email protected]‡MIT and NBER. E-mail: [email protected]
1 Introduction
Recent decades have seen substantial reductions in the barriers that impede trade be-tween nations—a process commonly referred to as ‘globalization’. But trade does not startor stop at national borders. The trading frictions faced by many households, especiallythose in developing countries, include not only the international trade costs that havefallen in recent times, but also the intra-national trade costs that separate these house-holds from their nearest port or border. The goal of this paper is to use newly collecteddata to provide an answer to an important and primitive empirical question: How largeare the costs that separate households in developing countries from the global economy?
Like a voluminous existing literature (reviewed below), we seek an answer to thisquestion by drawing inferences from the equilibrium distribution of prices over space—prices that are observed monthly and over several decades, for hundreds of consumerproducts, across thousands of local markets. in a sample of representative developingcountries.1 But unlike this literature we use new data and new tools to overcome threewell-known challenges that plague such inferences:
1. Spatial price gaps may reflect differences in unobserved product characteristics (such asquality) across locations. The overriding principle guiding our data collection effortsand resulting sample is to use exclusively products that are extremely narrowlydefined (analogously to barcodes).
2. Spatial price gaps only rarely are directly informative of trade costs. A standard argu-ment maintains that: if trading (ie arbitrage) is perfectly competitive, bilateral pricegaps for any pair of locations always place a weak lower bound on the cost of trad-ing between that pair of locations; and further, this inequality is binding if trade isactually occurring among the pair. Unfortunately, in practice the (highly detailed,intra-national) trade data required to apply this argument is unavailable. We havetherefore collected unique data on the precise source location of production (in thecase of domestically produced goods) or importation (in the case of imported goods)for each product in our sample. As we demonstrate below, without this informationour estimates of intranational trade costs would be biased downwards by a factorof almost two.
3. Spatial price gaps may reflect variable mark-ups across locations, not just the marginal costof trading. Put simply, arbitrage may not be perfectly competitive (ie arbitrage may
1Our current sample comprises Ethiopia and Nigeria and in ongoing work we are extending this sampleto include 5-10 additional developing countries.
1
not be free). If traders possess market power then spatial price gaps will be contam-inated by variation in mark-ups across locations. However, we demonstrate thatestimates of pass-through—which we estimate nonparametrically for each locationand product in our sample—can be used to correct for this contamination. The keyinsight is that estimated pass-through is a sufficient statistic for the extent to whichmark-ups vary across locations. Using our pass-through-based correction the es-timated marginal cost of trading is substantial, approximately twice as large as itwould be under the assumption of perfectly competitive trading (an assumptionthat is rejected by our data).
In short, we find that total intranational trade costs are large, and that a substantial com-ponent of inter-spatial price variation is due to varying levels of mark-ups that increas-ingly remote, uncompetitive locations pay for the delivery of goods produced abroad orafar.2
Our analysis has two important implications for how remoteness affects consumersand intermediaries situated in the interior of developing countries. The marginal cost ofdistance in our sample is extremely high, approximately seven to 15 times larger than themarginal cost of distance found by Hummels (2001) for truck shipments between Canadaand the US. The ability to consume goods at locations other than the port or the factorycreates a (partial equilibrium, that is holding prices in other markets contant) social sur-plus. These high marginal costs of distance imply that smaller quantities of this socialsurplus are available to consumers and intermediaries in remote locations (compared tothose in less remote locations).
In addition, because our estimates trace out how mark-ups vary over space we areable to speak to the share of surplus on traded goods that accrues to intermediaries (andto deadweight loss) in remote locations. A natural approach to measuring the distribu-tion of surplus along these lines would involve estimating mark-ups but in our devel-oping country context data on consumption quantities (for narrowly defined goods) istypically unavailable; standard methods for computing mark-ups (as described in, eg,Pakes (2008)) are therefore not possible. Fortunately—by an extension of the analysis inWeyl and Fabinger (2011)—if we restrict attention to demand systems that take a partic-ular form in which pass-through is constant (and hence CES demand is a special case)the rate of equilibrium pass-through (which we estimate for each location and product in
2Two important caveats, however, arise due to the nature of our restriction to a sample of only extremelynarrowly defined consumer products (a restriction which, to us, seems unavoidable in any attempt to mea-sure trade costs with price dispersion). First, these goods may not be representative of the entire nationalconsumption basket in our sample countries. And second, it is plausible that consumer products that canbe identified, essentially, at the unique barcode level are distributed differently than other products.
2
Step 2 above) is a sufficient statistic for the distribution of surplus, so estimates of mark-ups are unnecessary. Using this result we find that remote markets are less competitive,and that whatever social surplus from trade exists in remote locations sees larger sharesaccruing to intermediaries and DWL (relative to consumers).
The work in this paper relates to a number of different literatures. Most relevant isa recent and voluminous literature uses aspects of spatial price dispersion in order toidentify trade costs.3 Various segments of the literature have dealt with each of thesethree challenges in isolation, but we believe that our work is unique in attempting tocircumvent all three of these challenges. We discuss the response to these three challengesin the existing literature here in turn.
First, a large strand of this literature argues that inter-spatial arbitrage is free to enterand hence that inter-spatial price gaps place lower bounds on the marginal costs of trade,where these lower bounds are binding among pairs of locations that do trade. See, for in-stance, Eaton and Kortum (2002), Donaldson (2011), Simonovska (2010) and Simonovskaand Waugh (2011a), as well as the work surveyed in Fackler and Goodwin (2001) and An-derson and van Wincoop (2004).4 A central obstacle in this literature has been the need towork with narrowly defined products and yet also know which location pairs are actu-ally trading those narrowly defined products; our approach exploits unique data on thelocation of production of each product in our sample.
A second and recent strand of the literature draws on proprietary retailer or consumerscanner datasets from the US and Canada in order to compare prices of extremely nar-rowly identified goods (that is, goods with unique barcodes) across space (see for exampleBroda and Weinstein (2008), Burstein and Jaimovich (2009) and Li, Gopinath, Gourinchas,and Hsieh (2011)). However, this work typically lacks information on the region or coun-try of origin so point-identifying the level of trade costs is typically not the focus.5
3Another body of work to which our study relates is the rapidly growing literature on intermediationin trade, including (Ahn, Khandelwal, and Wei, 2011; Antras and Costinot, 2011; Bardhan, Mookherjee, andTsumagari, 2011; Chau, Goto, and Kanbur, 2009). This work aims to understand when trade is conductedvia intermediaries rather than by producers directly. Our work is instead focused on the consequencesof intermediaries—who potentially possess market power—for the magnitude of intranational barriers totrade, the pass-through of world price changes, and the distribution of the gains from trade.
4An additional body of work, (see for example Engel and Rogers (1996), Parsley and Wei (2001), Brodaand Weinstein (2008) and Keller and Shiue (2007)) uses moments derived from inter-spatial price gaps toinfer trade costs without data on which location pairs are actually trading. Because these moments pooltogether information from location pairs that are trading (on which price gaps are equal to trade costs) andlocation pairs that are not trading (for which price gaps understate trade costs) it is not clear how thesemoments estimate the level of trade costs without knowledge of the relative proportions of trading andnon-trading pairs in the sample.
5Li, Gopinath, Gourinchas, and Hsieh (2011), however, point out that their price gap estimates forlocation pairs on either side of the Canada-US border do place a lower bound on the cost of trading acrossthis border.
3
Finally, a third strand of this literature considers, as we do, the possibility that pro-ducers or intermediaries have market power (that is, arbitrage is not free to enter) andhence that firms may price to market. (See, for example, Feenstra (1989), Goldberg andKnetter (1997), Goldberg and Hellerstein (2008), Nakamura and Zerom (2010), Fitzgeraldand Haller (2010), Li, Gopinath, Gourinchas, and Hsieh (2011), Burstein and Jaimovich(2009), Atkeson and Burstein (2008), Alessandria and Kaboski (2011), and Berman, Mar-tin, and Mayer (2012)). In particular, this literature has placed heavy emphasis on theextent of exchange rate pass-through and its implications for market power. We apply asimilar logic to the market for each good and location in our sample with the goal beingto infer how intermediaries’ market power and equilibrium mark-ups vary across loca-tions, as well as how variable mark-ups over space cloud inference of how the marginalcosts of trading vary over space. In this sense, the paper is related to a recent literaturethat explores the interaction between the gains from trade and variable markups. (See,for example, Arkolakis, Costinot, Donaldson, and Rodriguez-Clare (2012), De Loecker,Goldberg, Khandelwal, and Pavcnik (2012), Edmond, Midrigan, and Xu (2011), Feenstraand Weinstein (2010) and Melitz and Ottaviano (2008)).
The remainder of this paper proceeds as follows. Section 2 describes the new datasetthat we have constructed for the purposes of measuring and understanding intranationaltrade costs in our sample of developing countries. Section 3 outlines a theoretical frame-work in which intranational trade is carried out by intermediaries who potentially enjoymarket power, as well as how we use this framework to inform empirical work that aimsto estimate the size of intranational trade costs as well as the distribution of the gainsfrom trade between consumers and intermediaries. Section 4 discusses the empirical im-plementation of this methodology and presents our findings. Section 5 concludes.
2 Data
Our study draws on two main sources of data: (i) the retail prices of products atvarious points in space (within developing countries), in time, and for many narrowly-defined products; and (ii) the location(s) of production or import of each product in oursample. We describe these two types of data here in turn.
2.1 Data on retail prices
The key requirement for studying the extent to which households in developing coun-tries are integrated with global markets is high quality price data. The methodology we
4
propose below requires observations of retail prices prevailing at many points in time,across many geographically segmented markets, for extremely narrowly defined prod-ucts (such that within-product differences in quality over space can be presumed to besmall). Fortunately the national statistical agencies of many developing countries collectexactly such data in the process of compiling a consumer price index. The data collectionexercise underpinning this paper has involved a search for a set of countries that collectdata that meets these standards and are willing to share their raw data (rather than thetypically publicly available aggregates) with researchers.
The typical country in our sample conducts a monthly price survey across tens of typ-ically fixed locations throughout the country, usually with representative regional cover-age. Enumerators are asked to survey particular retail establishments (the type of whichis often also available in the data that they share with us) and write down the postedprice (or typical sale price if a posted price is not available) for a fixed set of very nar-rowly defined products, and to record no observation if the product is not for sale atthe enumeration location on the enumeration date. Because the product list is typicallychosen to provide wide coverage of the a typical consumption basket many of the prod-ucts surveyed—such as rice, bread or haircuts—are not narrowly defined and we excludethese products from our analysis. We instead work with a sample of consumer productsthat are uniquely identified by their product descriptions. These descriptions include theproduct type, brand name, and specific size—such that the descriptions are akin to bar-codes that uniquely identify products in consumer scanner datasets in developed coun-tries. (Many countries that have made their data available to us contain product descrip-tions that lack unique, brand-level product identifiers and so these countries are excludedfrom our study.)
In the current version of this paper we use the following sub-samples of data only:
• Ethiopia (2001-2010): Covers 103 towns, the locations of which are illustrated in themap in Figure 1. Fifteen products are covered, which are detailed in Table 1.
• Nigeria (2001-2010): Covers 36 town markets (one per state), the locations of whichare illustrated in the map in Figure 2. Seventeen products are covered, which aredetailed in Table 2.
Our data collection efforts have lead us to the following additional sources of data (thoughthese data are not ready for analysis at present):
• Philippines (2000-2010): Covers all the provinces and major cities of the country (89unique geographic points in total). The data span 2000 to 2010 for close to 90 nar-rowly defined goods.
5
• India (1985-2010): We work with the raw data used to compute the rural CPI only.This data covers over 650 villages distributed across the 28 states of India and in-cludes price data on over 100 narrowly identified products. Some of these goodsvary by state and region, while others are present across the various states. For theperiod from 1985-1993 we obtained this data from archival records in the Indian BLSand from 1994-2010 we obtained this data from the NSSO.
• Zambia (1996-2005): Covers 48 district centers and 60-70 narrowly identified prod-ucts. This dataset also includes information on the typical quantity of sale, and thecharacteristics of the product (weight, volume, etc.).
• Bangladesh (2004-2010): Covers locations in all 64 districts of the country, each ofwhich is further divided into an urban and rural region. The number of commodi-ties in the data vary between rural and urban centers, from 30-50 narrowly identi-fied products. In addition, the dataset contains information on the type of retailerat which the price observation was made, as well as the weight and quantity of atypical unit of the good sold.
• Rwanda (2009-2011): Covers 48 towns over 5 regions (4 of which are further splitinto urban and rural centers) and 60 narrowly identified products.
• Senegal (2006-2010): Covers over 5 town markets and 20-30 narrowly identifiedproducts. As with the Nigerian data (from 2007-2010), the dataset includes infor-mation on the type of retailer at which a particular price observation was noted.
• In addition to the countries noted above, we have acquired price data on Guinea-Bissau and efforts are underway to acquire similar price data from other countriesincluding Pakistan, Mexico and Ghana.
2.2 Data on production source (factory or import) locations
For the case of domestically-produced products, we have conducted a telephone in-terview with the firms that produce each of the products in our sample. We ask each firmfor the location(s) of production that serve markets in each country in our sample, andask this information retrospectively so as to cover all of the years in our sample. We havealso sought to corroborate this information by surveying distributors.
For the case of imported products we have contacted distributors to learn the portof entry of each imported product in each country (and year) in our sample. Again, wecorroborate this information with official trade statistics data where possible.
6
3 Theoretical Framework
In this section we first describe a model of intranational trade carried out by interme-diaries who potentially enjoy market power. We then go on to discuss how this frame-work can be used to inform empirical work that aims to estimate the size of intranationaltrade costs as well as the distribution of the gains from trade between consumers andintermediaries.
3.1 Model Environment
We consider an environment in which there are D isolated locations6 indexed by dand K products indexed by k. Products k can be either domestically produced or importedfrom abroad. Domestically produced products are made at a factory location indexed by oand imported products are imported into the country through a port or border crossing atlocation o; regardless of whether the products are made at home or abroad, the domestic‘origin’ of the product is location o. (Note that we use the mnemonic o for origin and dfor destination.) We treat the market for product k traded from location o to location d asan isolated market—that is, we abstract for now from general equilibrium considerationsthat would introduce interactions in product or factor markets across or within locations.
We assume that any product k is sold on wholesale markets at the source (ie factorygate or port) location o for a price Pk
ot at date t. This product is then bought at the originlocation o wholesale market and traded from location o to any destination location d bydomestic intermediaries. These intermediaries specialize in the activity of purchasing aproduct in bulk at a wholesale market, transport the product to a destination location,and then finally selling the product to consumers at that location.7
Intermediaries incur costs of trading. Each intermediary has a total cost function,C(qk
odt), which is the cost of trading qkodt units of product k from location o to location d
at date t. Total costs are the sum of fixed costs of entry into the distribution sector, Fkodt,
and per-unit costs of trading. Per-unit costs are themselves the sum of the cost of buyingthe product at the origin location (which is simply the origin price, Pk
ot) and the marginalcosts of trading, denoted by τ = τ(Xk
odt). We assume (in Assumption 1) that marginalcosts are ‘specific’ (ie charged per unit of product shipped) and constant; future work will
6Locations are isolated in the sense that consumers do not travel to economies other than their own topurchase items. More generally, we simply require that intermediaries’ marginal costs are sufficiently low(relative to consumers’ travel costs) that consumer always buy goods locally from an intermediary ratherthan traveling themselves to other locations to make their purchases.
7Note that we assume that there is just one integrated sector that intermediates trade between producers(or importers) and final consumers, combining distribution and retail into one activity.
7
explore extensions to this basic case. Finally we let Xkodt denotes a set of marginal cost
shifters specific to the route from origin o to destination d (such as the quality of roadsalong the route) and the product k shipped.
Assumption 1. The cost to an intermediary of buying qkodt units of product k from location o at
date t (for an origin price Pkot) is given by the sum of fixed and (constant, specific) marginal costs:
C(qkodt) =
[Pk
ot + τ(Xkodt)]
qkodt + Fk
odt.
Intermediaries maximize profits by choosing the amount of the product to sell, qkdt.
Let Qkdt denote the total amount sold to the market by all intermediaries. The essential
strategic interaction across intermediaries is the extent to which an intermediary’s ac-tions (his quantity choice, qk
dt) affect other intermediaries’ profits through the aggregatequantity Qk
dt. We follow the ‘conjectural variations’ approach to oligopolistic interactions(e.g. Seade, 1980) and assume (in Assumption 2) that this relationship is summarized by
the parameter θkdt
.=
dQkdt
dqkdt
. The case of symmetric Cournot oligopoly corresponds to θ = 1,
the case of a pure monopolist corresponds also to θ = 1, while perfect competition corre-sponds to θ = 0. In what follows we will not take a stand on the value of θk
dt because ourempirical application will not need—nor be able—to identify this model parameter.
Assumption 2. Intermediaries selling product k in location d at date t perceive the effect of their
sales decision qkdt on aggregate sales Qk
dt to be given by the parameter θkdt
.=
dQkdt
dqkdt
that is fixed in
any period. Alternative values of this parameter span a range of market structure assumptions.
Given the above notation, and denoting consumers’ aggregate inverse demand curveby Pk
dt(Qkdt), each intermediary’s first order condition for profit maximization implies that:
Pkdt =
[Pk
ot + τ(Xkodt)]− θk
dt∂Pk
dt(Qkdt)
∂Qkdt
qkdt. (1)
As is the case for any producer, the price that intermediaries’ charge here (ie Pkdt) is equal
to the intermediaries’ total marginal costs (the sum of the purchase price at the origin, Pkot,
and the marginal cost of trading, τ(Xkodt)) plus the markup that intermediaries potentially
charge (which we denote by µkdt
.= −θk
dt∂Pk
dt(Qkdt)
∂Qkdt
qkdt).
It remains to specify the process of entry into the intermediary activity. The stock ofpotential intermediaries may potentially be constrained by credit constraints, reputationissues, caste or ethnic traditions etc. For this reason we assume (in Assumption 3) thatall intermediaries are identical and that the stock of intermediaries buying product k at
8
location o and selling it at location d on date t, denoted by mkodt, is fixed and exogenous
within any period.
Assumption 3. The number of identical intermediaries trading product k from location o tolocation d on date t, denoted by mk
odt, is fixed and exogenous.
Below we make extensive use of the concept of pass-through. This is definedas the amount by which intermediaries’ equilibrium prices respond to a change in their
marginal costs; that is, we define pass-through as ρkodt
.=
dPkdt
dPkot=
dPkdt
dτ(Xkodt)
. Differentiating
equation (1) it can be shown that, in general, pass-through takes the form
ρkodt =
[1 +
(1 + Ekdt)
φkodt
]−1
, (2)
where we define φkodt
.=
mkodt
θkdt
as the ‘competitiveness index’ (since it rises in the number
of intermediaries, mkodt, and falls in these intermediaries’ perceived individual influence
on aggregate supply, θkdt) and Ek
dt.=
Qkdt(
∂Pkdt
∂Qkdt
) ∂
(∂Pk
dt∂Qk
dt
)∂Qk
dtQ 0 is the elasticity of the slope of
inverse demand. (Note that Ekdt = 1
ekdt− 1 − Qk
dtek
dt
∂ekdt
∂Qkdt
, where ekdt
.=
∂Qkdt
∂Pkdt
Pkdt
Qkdt≤ 0 is the
elasticity of demand.) As this expression makes clear, pass-through depends on only twomarket characteristics: the competitiveness (φk
odt) of the market (where importantly it is
only φkodt
.=
mkodt
θkdt
that matters, not mkodt or θk
dt individually) and the second-order curvature
of the demand curve (ie Ekdt, the elasticity of the slope of demand).
The above results hold for any demand curve (or, more generally, to the single-itemconditional demand relationships in any demand system). However, in order to simplifya number of results below we will at times make the additional assumption (in Assump-tion 4) that consumer preferences belong to a particular class of demand for which Ek
dtis constant. We refer to this as ‘constant pass-through demand’ (though note that equi-librium pass-through, ρk
odt would only be constant under this demand class if the com-petitiveness index, φk
odt, were also constant). Constant pass-through demand was firstidentified by Bulow and Pfleiderer (1983) and is a natural generalization of isoelastic de-mand. Indeed, Bulow and Pfleiderer (1983) prove that the only demand system withconstant pass-through is the class introduced here.
Assumption 4. Consumer preferences take the constant pass-through demand form such that
9
total demand Qkdt depends on price Pk
dt in the following manner:
Qkdt(Pk
dt) =
(
akdt−Pk
dtbk
dt)
1δkdt if (Pk
dt ≤ akdt, bk
dt > 0and δkdt > 0)or (Pk
dt > akdt, bk
dt < 0and δkdt < 0)
0 if Pkdt > ak
dt, bkdt > 0and δk
dt > 0
∞ if Pkdt ≤ ak
dt, bkdt < 0and δk
dt < 0
with akdt ≥ 0. Accordingly, inverse demand is:
Pkdt(Q
kdt) = ak
dt − bkdt
(Qk
dt
)δkdt . (3)
For this demand system we have ekdt = − 1
δkdt
(Pk
dtak
dt−Pkdt
)≤ 0 and Ek
dt = δkdt − 1Q 0;
that is, by design, Ekdt is equal to a (constant) model parameter, but this parameter is
free to vary. Note that the case of isoelastic demand corresponds to a restriction of thisdemand class in which ak
dt = 0. Hence from equation (2) equilibrium pass-through underAssumption 4 is equal to
ρkodt =
[1 +
δkdt
φkodt
]−1
(4)
Equilibrium pass-through can be ‘incomplete’ (ie ρkodt < 1) for δk
dt > 0 and ‘morethan complete’ (ie ρk
odt > 1) with δkdt < 0. Hence nothing in this class of preferences
restricts whether pass-through will rise or fall with the remoteness of locations within acountry; the only restriction is that pass-through is constant. Finally, note that, whateverthe demand parameter, the state of competitiveness (summarized by φk
odt) matters forequilibrium pass-through; in particular, if competition were perfect (ie φk
odt → ∞) thenequilibrium pass-through is ‘complete’ (ie ρk
odt = 1) for any demand parameters.
3.2 Using the Model to Measure Intranational Trade Costs
In what follows our goal is to describe how the theoretical framework introducedabove can be used, in conjunction with the data described in Section 2 above, to estimatethe magnitude of intranational trade costs. Additionally, we describe how our theoreticalframework can be used to estimate the distribution of surplus (among consumers andintermediaries) for each location in our sample. We break down our analysis into threesteps as follows.
10
3.2.1 Step 1: Using price gaps to measure total intranational trade costs
We define total intranational trade costs—denoted here by Tkodt—in a manner that we
feel is relevant from the perspective of final consumers: total trade costs are the price thata final goods consumer must pay for an intermediary to deliver a good from its originlocation to the consumer’s location (the intermediary’s destination). Equation (1) abovedescribes how in this framework intermediaries are potentially charging a mark-up ontrades such that total trade costs are the sum of intermediaries’ marginal costs and inter-mediaries’ mark-ups. The following result is then immediate.
Result 1: Under Assumption 1 the total cost, Tkodt, of trading product k from its origin loca-
tion o to its destination location d at date t is equal to the price of this product across those twolocations on that date (ie Pk
dt− Pkot). Further, this total trade cost is equal to the sum of the marginal
costs of trading, τ(Xkodt), and the mark-up charged by intermediaries, µk
dt. That is:
Pkdt − Pk
ot = Tkodt
.= τ(Xk
odt) + µkdt. (5)
However, the difference in prices for product k among two distinct destination locations, i and jfor i 6= o and j 6= o, is uninformative about the total cost of trading among those locations. Thatis:
Pkjt − Pk
it R Tkijt for i 6= o and j 6= o. (6)
Result 1 is extremely simple, yet it offers a powerful guide to empirical work. Armedwith a dataset of prices prevailing for a given product at several locations, Result 1 sug-gests which price gaps over pairs of locations are informative of total trade costs andwhich are not. This is important because most researchers do not observe the origin loca-tion of each product in each time period, so they do not know which location(s) in theirdataset, if any, correspond to the origin location o, which is required to apply Result 1.In Section 4.1 below we use unique data on the production/importation location(s) ofeach product and year in our dataset and thereby apply Result 1 in order to estimate themagnitude of intranational trade costs for a group of developing countries. We also dis-cuss the size of the bias one would obtain in our dataset without knowledge of originlocations.
11
3.2.2 Step 2: Estimating pass-through rates
As discussed in the Introduction, our method for inferring the extent of mark-up vari-ation over space (a necessary input into estimating the marginal costs of intranationaltrade) relies on flexible estimates of the extent of equilibrium pass-through for each prod-uct in each location. We discuss here how the theoretical framework outlined above canbe used to provide these estimates. As long as exogenous variation in the origin (or bor-der) price can be isolated, a reduced-form pass-through parameter can be estimated foreach product k, destination location d and time window (noting that at least two dates twould be required to study how a change in the origin price affects destination prices).
It is straightforward to show, using Equation (4) above, the following result, whichcharacterizes the relationship between destination and origin prices—that is, equilibriumpass-through—as well as the relationship between pass-through and underlying struc-tural parameters (ie mk
odt, θkdt, and δk
dt).
Result 2: Under Assumptions 1-4 the relationship between destination prices Pkdt and origin
prices Pkot for any product k and at any date t satisfies
Pkdt = ρk
odtPkot + ρk
odtτ(Xkodt) + (1− ρk
odt)akdt, (7)
where ρkodt =
(1 + δk
dtφk
odt
)−1
and φkodt
.=
mkodt
θkdt
.
This implies that a regression of destination prices (Pkdt) on origin prices (Pk
ot), con-ditional on controls for both the marginal cost of trading (ie τ(Xk
odt)) and local demandshifters (ie ak
dt), would reveal the equilibrium pass-through rate (ρkodt) inherent to each des-
tination market and product. Unfortunately, both the marginal cost of trading and localdemand shifters are unobservable to researchers—indeed, if these were observable thenanswers to the questions we pose in this paper would be immediately available. Nev-ertheless, in Section 4.2 below we propose an empirical strategy that aims to control forthese variables and hence provide consistent estimates of the equilibrium pass-throughrate (ρk
odt) prevailing in each destination location d and product k separately. While theprincipal reason for obtaining these estimates is, as described in the next sub-section, toinfer how mark-ups vary over space, Result 2 demonstrates that an additional use for
pass-through estimates is to estimate φkodt via the formulaρk
odt =
(1 + δk
dtφk
odt
)−1
.
12
3.2.3 Step 3: Using pass-through estimates to estimate the marginal costs of distance
A central aim of much of the literature on the estimation of trade costs has been tounderstand the determinants of the marginal costs of trading. In a setting of perfect com-petition, total trade costs—as we have defined them above—are equal to the marginalcosts of trading since mark-ups are zero. In such a setting the simple price gap methodsdescribed above in Step 1 are sufficient for determining the magnitude and determinantsof the marginal costs of trading.
However, once it is possible that intermediaries enjoy market power, price gaps reflectnot just the marginal costs of trading over space but also the mark-ups that intermediariescharge when carrying out trades over space. If these mark-ups vary over space (for ex-ample because the state of competition in the intermediary sector varies over space, orsimply because preferences are such that pass-through is anything but complete), thenthe change in price gaps over space reflects not just how space imposes marginal costsbut also how mark-ups vary over space.
The central challenge here is to separate out these two effects without the ability to sep-arately estimate marginal costs and mark-ups (a challenging prospect in any setting, butone that is especially challenging here where researchers lack access to data on consumerquantities or to producers’ input choices). In this section we describe a method—whichdraws on the estimates of equilibrium pass-through ρk
od obtained in Step 2 above—thataddresses this challenge. The key intuition is that pass-through describes how pricesrespond to any marginal cost shock. We have described above a way to estimate pass-through by using the observed response of destination prices to origin price shocks. Wenow use these pass-through estimates to deduce the extent to which the marginal costsof trading affect prices (which, by definition, they do via the extent of pass-through) andhence infer the true marginal costs of trading over space from observed price variationover space. This logic is formalized in Result 3 below which is a direct implication ofResult 2 above.
Result 3: Under Assumptions 1-4 the relationship between ‘adjusted price gaps’ between ori-
gin and destination locations for any product k and at any date t (ie Pkdt−ρk
odtPkot
ρkodt
) and the marginal
costs of trading that product between those locations on that date (ie τ(Xkodt)) satisfies:
Pkdt − ρk
odtPkot
ρkodt
= τ(Xkodt) +
(1− ρkodt)
ρkodt
akdt. (8)
13
Recall from Result 1 above that origin-destination price gaps (ie Pkdt − Pk
ot) are equal tothe sum of marginal costs of trade (ie τ(Xk
odt)) and mark-ups charged by intermediaries;because mark-ups potentially vary with distance, the extent to which price gaps vary overspace does not identify the extent to which the marginal costs of trading vary over space.
By contrast, a key message of Result 3 here is that ‘adjusted price gaps’ (ie Pkdt−ρk
odtPkot
ρkodt
)
are equal to sum of the marginal costs of trade (ie τ(Xkodt)) and a pass-through adjusted
demand shifter (ie (1−ρkodt)
ρkodt
akdt). This suggests that, with suitable controls for this demand-
shifter, the extent to which adjusted price gaps—rather than simple price gaps—vary overspace does identify the extent to which the marginal costs of trading vary over space. InSection 4.3 below we apply this method directly in order to estimate how distance (ie avariable Xk
odt that shifts τ(Xkodt)) affects the marginal costs of trading.
3.3 Using the Model to Measure Division of Surplus
In any market setting where producers enjoy market power a natural question con-cerns the share of surplus accruing to these producers, as well as the share of surplusthat is destroyed (ie deadweight loss) due to their market power. The setting we considerhere—in which intranational trade between producers and consumers is carried out byintermediaries who potentially enjoy market power—is no exception. In this setting thesurplus in question is essentially the gains from trade. That is, consumers in destinationlocation d benefit from being able to consume product k sourced from the origin location o(be that origin a domestic factory or a port through which foreign producers’ goods enter)because there are gains from trade (ie the product would cost more to produce in locationd).
Naturally it is challenging to identify the share of surplus accruing to consumers, in-termediaries and deadweight loss. And this is especially challenging in settings like ourswhere researchers lack data on the quantities of narrowly defined products consumed(which could in principle be used to estimate demand curves, mark-ups and hence thedivision of surplus). Fortunately, based on an extension of the logic in Weyl and Fabinger(2011), in the theoretical framework we have outlined above there exists a simple connec-tion between pass-through and the division of surplus, which allows an answer to answerthis question without the need for data on quantities consumed.
To see this, consider first the calculation of the amount of consumer surplus generatedby any partial equilibrium market setting (that is, where the prices in all other marketsare held constant) for product k in destination market d at date t. Consumer surplus
when Qkdt is supplied to the market is defined as CSk
dt(Qkdt)
.=∫ Qk
dtψ=0[P
kdt(ψ)− Pk
dt(Qkdt)]dψ,
14
where Pkdt(ψ) is the consumers’ inverse demand curve evaluated at argument ψ and
Qkdt is the total amount consumed in equilibrium in the market. Since dCSk
dt(Qkdt(ψ))
dPkot
=
−Qkdt(ψ)
dPkdt(Q
kdt(ψ))
dQkdt
dQkdt
dPkot
consumer surplus can also be written in a way that stresses its
essential connection with pass-through:
CSkdt(Q
kdt) = −
∫ ∞
ψ=Pkot
Qkdt(ψ)ρ
kodt(ψ)dψ. (9)
Following similar steps we now calculate the amount of surplus captured by inter-mediaries in this setting. Intermediaries’ surplus when Qk
dt is supplied to the market isdefined as total variable profits among intermediaries, or ISk
dt(Qkdt)
.= mk
odtΠkodt(q
kodt) =
mkodt
∫ ∞ψ=Pk
ot
dΠkodt(ψ)
dPkot
dψ.8 Differentiating total profits we have:
dΠkodt(ψ)
dPkodt
=(mk
odt − θkdt)q
kodt(ψ)
(mkodt + θk
dt) + θkdtE
kdt(ψ)
− qkodt(ψ),
=
(φk
odt − 1
φkodt
)ρk
dt(ψ)qkodt(ψ)− qk
odt(ψ), (10)
where, recall, Ekdt(ψ) is the elasticity of the slope of demand and ρk
odt(ψ) =
[1 + (1+Ek
dt(ψ))
φkodt
]−1
when each is evaluated at the argument ψ, andφkodt
.=
mkodt
θkdt
is the ‘competitiveness index’
first introduced above. This result holds for any demand curve. Using this result, inter-mediaries’ surplus can be written as:
ISkdt(Q
kdt) = −
∫ ∞
ψ=Pkot
Qkdt(ψ)dψ +
(φk
odt − 1
φkodt
) ∫ ∞
ψ=Pkot
Qkdt(ψ)ρ
kodt(ψ)dψ. (11)
Similar logic implies that intermediaries’ surplus can also be written as:
ISkdt(Q
kdt) = −
∫ Pkdt+τ(Xk
odt)
ψ=Pkdt
µkdt(ψ)dψ +
(φk
odt − 1
φkodt
) ∫ Pkdt+τ(Xk
odt)
ψ=Pkdt
µkdt(ψ)ρ
kodt(ψ)dψ, (12)
where µkdt(ψ) is the mark-up of intermediaries evaluated at the argument ψ. As expected,
8We work with a notion of surplus defined on total variable profits in part because nothing in ourdataset can be used to estimate the fixed costs intermediaries pay. While this overstates intermediaries’total profits it does not overstate consumer surplus or deadweight loss since the fixed costs of productionconsume resources available to society.
15
under the perfectly competitive limit (ie where the competitiveness index φkodt → ∞ and
hence, by equation (2), ρkodt(ψ)→ 1 for any value of ψ) there is no intermediaries’ surplus
(ie ISkdt(Q
kdt)→ 0 for any value of Qk
dt).Finally, following similar steps we calculate the amount of deadweight loss—denoted
DWLkdt(Q
kdt) when Qk
dt is supplied to the market—created by market power in this setting.Using similar arguments to those above this can be written as
DWLkdt(Q
kdt) = −
∫ Pkdt+τ(Xk
odt)
ψ=Pkdt
µkdt(ψ)ρ
kodt(ψ)dψ, (13)
where again µkdt(ψ) is the mark-up of intermediaries evaluated at the argument ψ. Natu-
rally, under the perfectly competitive limit (where mark-ups are always zero) this tendsto zero.
Applying equations (9), (11), (12) and (12) it is then straightforward to show the fol-lowing result:
Result 4(a): Under Assumptions 1-3, the ratio of intermediaries’ surplus ISkdt(Q
kdt) to consumer
surplus CSkdt(Q
kdt) in the market at destination location d for product k on date t is given by
ISkdt
CSkdt
(Qk
dt
)=
1(ρQ)k
dt
+1− φk
odt
φkodt
, (14)
where(ρQ)k
dt is a quantity weighted average of the pass-through rate, defined as
(ρQ)k
dt.=
∫ ∞ψ=Pk
otQk
dt(ψ)ρkodt(ψ)dψ∫ ∞
ψ=Pkot
Qkdt(ψ)dψ
. (15)
Similarly, the ratio of deadweight loss DWLkdt(Q
kdt) to intermediaries’ surplus ISk
dt(Qkdt) in this
market is given byDWLk
dt
ISkdt
(Qk
dt
)=
(ρµ
)kdt
φkodt + (φk
odt + 1)(ρµ
)kdt
, (16)
where(ρµ
)kdtis a mark-up weighted average of the pass-through rate, defined as
(ρµ
)kdt
.=
∫ Pkdt+τ(Xk
odt)
ψ=Pkdt
µkdt(ψ)ρ
kodt(ψ)dψ∫ Pk
dt+τ(Xkodt)
ψ=Pkdt
µkdt(ψ)dψ
. (17)
16
These results, which are derived for a completely general demand structure, high-light the close connection between pass-through and the division of surplus in a generaloligopolistic setting. However, pass-through enters these formulae always as a weightedaverage of pass-through values between relative points (either between the equilibriumprice and an infinite price as in the case of
(ρQ)k
dt, or between the equilibrium price andmarginal cost as in the case of
(ρµ
)kdt). Unfortunately in our setting the weights in these
weighted average formulae (consumption quantities Qkdt(ψ) or mark-ups µk
dt(ψ) in(ρQ)k
dtand
(ρµ
)kdt, respectively) are not observed, nor is there any hope of credibly estimating the
demand structure so as to estimate these weights because consumption quantities are notobserved.
However, in the case of the constant pass-through class of demand (ie that describedin Assumption 4), pass-through (conditional on a fixed competitiveness index) is con-stant and hence weighted averages of this constant are equal to the constant; that is, theweights in Result 4(a) need not be observed. This statement is formalized in Result 4(b):
Result 4(b): Under Assumptions 1-4, the ratio of intermediaries’ surplus ISkdt(Q
kdt) to consumer
surplus CSkdt(Q
kdt) in the market at destination location d for product k on date t is given by
ISkdt
CSkdt
(Qk
dt
)=
1ρk
dt+
1− φkodt
φkodt
, (18)
where ρkdt is the constant equilibrium pass-through rate in this market and ρk
odt =
(1 + δk
dtφk
odt
)−1
.
Similarly, the ratio of deadweight loss DWLkdt(Q
kdt) to consumer surplus CSk
dt(Qkdt) in this market
is given byDWLk
dt
CSkdt
(Qk
dt
)=
φkodt + ρk
dt(φkodt − 1)
φkodt
[φk
odt − ρkdt(φ
kodt − 1)
] . (19)
Combining these two results, the share of total surplus accruing to consumers is given by
CSkdt
CSkdt + ISk
dt + DWLkdt
(Qk
dt
)=
[1
ρkdt+
1φk
odt+
φkodt + ρk
dt(φkodt − 1)
φkodt
[φk
odt − ρkdt(φ
kodt − 1)
]]−1
.
This result describes how, under Assumption 4 (ie a constant pass-through demandsystem), shares of surplus distributed in equilibrium among consumers, intermediaries’and deadweight loss in any market are all simple functions of just the equilibrium pass-
17
through rate ρkdt and the competitiveness index φk
odt prevailing in that market. Conditionalon obtaining estimates of ρk
dt and φkodt, therefore, Result 4(b) provides a direct estimate of
the division of surplus. We pursue this in Section 4.4 below.
3.4 Summary of Theoretical Framework
Using the methodology outlined above we can answer the question posed in the In-troduction: How large are intranational trade costs? To summarize, our answer to thisquestion is achieved in three steps:
1. Step 1: Use price gaps to measure total intranational trade costs. Among the pairs ofmarkets that are actually trading goods, that is between origin and destination mar-ket pairs, these trade costs (for any good, market pair and point in time) can beidentified simply as the price gap, Pk
dt − Pkot.
2. Step 2: Estimate pass-through rates. For each product k and destination market d weestimate a separate equilibrium pass-through rate ρk
od.
3. Step 3: Use pass-through adjusted price gaps to measure how distance affects the marginalcosts of intranational trade. The theory above has demonstrated how price gaps overspace consist of both marginal costs of trading and intermediaries’ mark-ups. How-ever, armed with estimates of pass-through ρk
od from Step 2 above, we have shownhow an ‘adjusted price gap’ formula does reveal how marginal costs of trade varywith shifters to thoses costs, such as distance. The intuition for this is straightfor-ward: pass-through embodies, by definition, how marginal costs affect equilibriumprices, so once estimates of pass-through are known (from time-series variation) thisinformation is vital for studying marginal costs through observations on prices.
In addition, our model highlights the close connection between pass-through rates andthe division of surplus in a partial equilibrium setting. That is, the pass-through estimatesobtained in Step 2 above provide sufficient statistics for the calculation of the relativeshares of social surplus accruing to intermediaries, to consumers, and to deadweight loss.
4 Empirical Results
4.1 Step 1: Using price gaps to measure total intranational trade costs
As described above, our first goal is to measure the magnitude of the total costs ofconducting intranational trade in developing countries. We define these total costs as
18
the price a consumer would have to pay to buy a good produced (or imported from) ata non-local source within her own country. Defined in this way, total intranational tradecosts reflect both the marginal costs of intranational trade and the potential mark-ups thatintermediaries’ charge on intranational trades. But regardless of their composition, thesetotal trade costs reflect the extent to which consumers pay additional costs to purchasenon-local goods.
Result 1 above described how, for a given commodity, total trade costs for any givendestination consumption location are then simply equal to the difference between theprice of that commodity at its source and the price of that same commodity at the desti-nation location. In the notation laid out above, we have
Pkdt − Pk
ot = τ(Xkodt) +µk
dt = Tkodt,
where Tkodt is the total cost of trading good k from location o to location d at time t. By
definition this is the difference in price between that at the destination (Pkdt) and that at the
origin (Pkot). The total cost of trading is the sum of the marginal cost of trading (τ(Xk
odt)),and the mark-up charged by intermediaries (µk
dt).It is important to note that the logic of the above result is only valid for inference
drawn product by product, and on data for which it is reasonable to assume that observa-tions on product k at different points in space are effectively exactly the same product. Forthis reason we work exclusively with products that are extremely narrowly defined, withprecision similar to barcode level identifiers. While this affects the representativeness ofour basket of goods (in terms of consumption weights) it is hard to imagine pursuing anyother approach without confronting serious issues of unobserved quality variation overspace.
A second important feature of Result 1 is that the result that Pkdt − Pk
ot = Tkod, is only
true for pairs of locations, o and d, that are actually trading product k. Our theory saysnothing about the relationship between price gaps and trade costs for pairs of locationsthat are not trading. More generally, for any pair of locations i and j all that a theory ofarbitrage can say (without knowledge of whether the two locations are trading) is that
Pkit − Pk
jt ≤ Tkijt.
Hence, for pairs of locations for which it is not known that trade is occurring, price gapssay nothing about the actual magnitude of total trade costs—both zero and infinite totaltrade costs are consistent with any data set. It is for this reason that we proceed in thissection using information only on those pairs of locations that are actually trading each
19
particular commodity; all other pairs of locations are effectively uninformative of thecosts of trading.
While in principle one could use trade data to infer whether two given locations aretrading product k at time t, there are two serious practical obstacles in doing so. First,trade data is rarely available within countries (especially developing countries), and evenmore rarely with the spatial precision required to study location-to-location trade. Sec-ond, trade data are rarely available at the product (ie brand-name) level so there is rarelya chance to know whether product k is traded or not.
Our approach utilizes, in lieu of trade data, a simple approach to infer which locationpairs are actually trading each commodity: we simply infer the production (or import)location of each good (in each year) in our sample. While trade may not literally occurfrom a given source location, separately, to each destination location (for example, trademay follow a hub-and-spoke arrangement where the hub is a location in the center of acountry) our approach still identifies the trade cost along the route actually followed fromlocation o to location d. An important limitation of our approach, however, is that thereare many pairs of locations that are not trading (the goods in our sample) and betweenwhich we therefore cannot infer the costs of trading. For this reason we seek to estimatethe fundamental underlying relationship between trade cost shifters (such as distance)and trade costs, rather than the particular costs of trading among every possible pair ofmarkets.
Tables 1 and 2 describe (for Ethiopia and Nigeria respectively), by commodity, theaverage price gap among trading pairs, as well as the standard deviation of these pricegaps, the average source price, and the average distance to the source location. In orderto account for inflation over the sample period and different currency units, all prices areconverted into 2001 US dollars (using the base period exchange rate).9 As can be seen,trade costs (ie from the variable, ‘price gaps among trading pairs’) can be substantial bothin absolute units (2001 US dollars) and in relative terms (ie as a percentage of the priceof a good at its source). The tables also reports the average price gap among all pairs (ieboth trading and non-trading pairs), for comparison.
The variable that we refer to as ‘distance’ (both here and throughout the paper) isactually the calculated minimum travel time (using calculations performed by GoogleMaps), so it can be thought of as a metric for distance that is adjusted for road quality andthat allows traders to take (what Google Maps believes is) the quickest route from source
9The price index for each period is obtained by calculating the average across all goods of the propor-tional price changes at the good’s origin location. The normalized prices are converted into 2001 US dollarsusing the prevailing exchange rate during the first month of the sample.
20
to destination. We can infer the approximate algorithm used by Google in these twocountries by sampling travel times along various road types at a random set of locations.The average minutes per mile to travel along primary, secondary and tertiary roads ineach country are shown below.
Approximate Minutes/Mile (Google Maps)Road Quality Ethiopia Nigeria
National highway 1.2 1.2Secondary road 1.4 1.4Tertiary road 1.9 2.4
Figures 3 and 4 (again, for Ethiopia and Nigeria respectively) plot the extent to whichthese price gaps, among trading pairs only, co-vary with the (log) distance separating thetrading pairs (that is the source and destination location). Additionally, the figures plotthe extent to which absolute price gaps co-vary with the (log) distance separating the loca-tions for all (unique, non-trivial) pairs of locations. These are semi-parametric regressionsthat include product-time fixed effects but allow distance to enter non parametrically viaa local polynomial. In what follows we will only be able to identify the marginal costsof distance. Hence the plots are normalized such that the cost of distance is zero for allspecifications at the bottom of the range of observed distances. As these figures illustrate,on average there is an upward sloping relationship, implying that distance raises the totalcosts of intranational trade. The slope is approximately twice as steep when we restrictour sample to only trading pairs, a finding we will return to shortly.
An important lesson from the theory outlined above is that price gaps over space,among trading pairs, reflect both the marginal costs that traders face and the mark-upsthat they may charge if they possess market power. Because of this, the relationship ofprice gaps with distance (as reported nonparametrically in Figures 3 and 4) is a mixtureof how marginal costs vary with distance (presumably positively and monotonically) andhow mark-ups vary with distance (either because the gap between the choke price andcosts is higher in inland destinations or because of intermediaries’ market power varyingover space at destination locations further and further away from the origin location). Animportant goal of Step 3 below is a separation of these two interacting relationships withdistance in order to recover the true effect of distance on the marginal costs of trading.
Finally, Tables 3 and 4 report coefficient estimates from regressions in which we regresslocal currency price gaps between two locations on the (log) distance between the two lo-cations. Column (2) estimates this relationship on the sample of trading location pairs
21
only, while column (1) estimates this relationship on all (unique, non-trivial) pairs of lo-cations. The results in column (2) show that there is a large and statistically significantrelationship between price gaps and (log) distance. The estimated coefficient in column(2) corresponds to a rise in the price gap of 2.99 cents and 3.40 cents cents to ship a goodfor each additional unit of log-distance (measured in minutes of travel time) in Ethiopiaand Nigeria respectively.10 Remarkably, these estimates are similar across the two Africancountries in our sample yet nothing in our analysis imposed this.
Notably these estimates in column (2), which are obtained from trading location pairsalone, are considerably different than those in column (1) which use all pairs of locations.The estimates obtained in column (1) draw on variation that is informative about the costsof distance (ie trading pairs) and variation from pairs of locations that are not tradingand whose variation is therefore completely uninformative about the (point-identified, asopposed to set-identified) costs of distance. The finding that the estimated cost of distanceare smaller when the uninformative pairs are included is not surprising. The price gapbetween any non-trading pair i and j is a function of the difference in trade costs incurredtransporting goods from the origin to locations i and form the origin to location j. Thetriangle inequality implies that the distance between i and j will be weakly larger thanthe difference between the distance from o to i and the distance from j to i. Therefore,a regression of uninformative price gaps on the distance between i and j will tend tounderestimate the marginal costs of distance.
While the results in column (2) speak to how distance raises total trade costs theydo not, for reasons discussed above, indicate that distance necessarily poses significantphysical costs of trading. It could simply be the case that intermediaries delivering to in-creasingly remote locations charge higher mark-ups than do intermediaries at proximatelocations. Our analysis in Step 3 below aims to understand how much characteristicssuch as distance raise actual physical costs of trading (ie intermediaries’ marginal costs)by separating the price gap-to-distance relationship into its two constituent parts: themarginal cost-to-distance relationship and the mark-up-to-distance relationship.
4.2 Step 2: Estimating pass-through
In Step 3 below we aim to measure how the marginal cost of intranational trade riseswith distance. Doing so requires a method for differentiating the extent to which distanceaffects the marginal costs of trading from the extent to which distance affects the mark-
10Recall all prices are normalized to base period prices (ie January 2001) and converted to US dollarsusing the January 2001 exchange rates.
22
up charged by traders. Pass-through, defined as the extent to which a marginal costshock raises the equilibrium price (and is hence equal to one minus the extent to whicha marginal cost shock raises mark-ups), is an essential ingredient for this analysis. Thecurrent section—Step 2— aims therefore to provide estimates of pass-through for eachlocation d and product k in our sample.
Before constructing these estimates it is worth emphasizing that pass-through is anobject of policy interest in its own right. This is for two reasons. First, pass-throughmeasures the extent to which households in developing countries are exposed to changesin economic conditions beyond their borders. In a market economy these effects workthrough price signals. It is therefore important to understand the extent to which theprices faced by households in developing countries—and especially those in remote seg-ments of developing countries—are actually affected by foreign price developments (suchas tariff changes, exchange rate changes, improvement in international shipping technolo-gies, or fluctuations in world demand and supply). If the prices paid by remote house-holds (for imported goods) are largely unaffected by foreign price developments then thequestion of whether integration with world markets via border policies (such as tradeliberalization or exchange rate policy) has been good or bad for these households is anon-starter. Second, as discussed in Result 2 above, the extent of pass-through is also ametric for the extent to which the market in a location is perfectly competitive: perfectcompetition implies complete pass-through, while incomplete pass-through is prima facieevidence for imperfect competition. This logic applies equally to domestically producedgoods (ie the extent to which a shock to the price at the factory gate location affects eachdestination location’s retail price) and to imported goods (ie the extent to which a for-eign price development, such as an exchange rate appreciation, affects each destinationlocation’s retail price).
Recall from Result 2 above that pass-through (ρkodt) relates to the extent to which ex-
ogenous origin prices (Pkot) affect endogenous destination prices (Pk
dt) in the followingmanner:
Pkdt = ρk
odtPkot + ρk
odtτ(Xkodt) + (1− ρk
odt)akdt. (20)
When using this equation to estimate pass-through (ρkodt) three identification challenges
arise.First, there is no hope of estimating a separate pass-through rate ρk
odt for each timedestination d, product k and time period t. We therefore focus on estimating the averagepass-through rate (which we denote ρk
od) for a destination location d and product k acrossall time periods in our sample. Because the pass-through rate is always equal to ρk
odt =
23
(1 + δk
dtφk
odt
)−1
under A1-A4, the assumption that the pass-through rate ρkod is constant over
time (within a product and location) amounts to assuming that the second-order propertyof demand δk
dt (for a product in a location) and the competitiveness parameter φkodt (for the
sale of a product in a location) are constant over time.Second, estimation of ρk
od here requires controls for the marginal cost of trading (ieτ(Xk
odt)) and for local demand shifters (ie akdt). Unfortunately, both the marginal cost of
trading and local demand shifters are unobservable to researchers—indeed, if these wereobservable then answers to the question posed in this paper would be immediately avail-able. In the absence of such controls we make the weakest possible assumption requiredto identify ρk
od, namely that the product-specific variation in the marginal cost of tradingand local demand shifters within destinations over time is orthogonal to the variation inthe origin price over time (or at least to an instrument for changes in the origin price overtime). That is, we assume that the marginal cost of trading, τ(Xk
odt), can be decomposedinto local but time-invariant (βk
1od), local but trending (βk2odt), macro but time-varying (β3t)
and residual (ζkodt) factors as follows: τ(Xk
odt) = βk1od + βk
2odt + β3t + ζkodt. Analogously, we
assume that destination market additive demand shocks, akdt, from Equation (3) above can
be decomposed as follows: akdt = αk
1d + αk2dt + α3t + νk
dt. Note that while this assumptionplaces certain restrictions on how the additive demand shifter, ak
dt, varies across locations,time and products, we place no restrictions on the multiplicative demand shifter, bk
dt, fromEquation (3) above. Conditional on these assumptions we estimate pass-through rates ρk
odby location and product by estimating the following specification,
Pkdt = ρk
odPkot + γk
od + γkodt + γt + εk
dt, (21)
where Pkdt is the destination price, Pk
ot is the origin price, γkod and γt are product-destination
and time fixed-effects, respectively, γkodt is a product-destination linear time trend, and
εkdt = ρk
odζkodt + (1− ρk
od)νkodt is an unobserved error term.
Third, estimation of equation (21) via OLS requires the additional assumption thatE[Pk
otζkodt]= 0 and E
[Pk
otνkdt]= 0, namely that the origin price Pk
ot is not correlated withthe time-varying and local shocks to local (destination location d-specific) trade costs ordemand shifter. If origin prices are set abroad (in the case of imported goods), or arepinned down by production costs at the factory gate (in the case of domestic goods), orare set on the basis of demand shocks at the origin location (which we omit from ouranalysis), then this orthogonality restriction seems plausible. But a nation-wide demandshock for product k (note that a nation-wide demand shock for all goods is controlled forwith the γt fixed effect) would violate this assumption. In future versions of this paper
24
we aim to explore the plausibilty of this assumption by estimating equation (21) via aninstrumental variables method in which the IV for the origin price is the price of a produc-tion input sourced from abroad (indeed some products k are produced entirely abroad)or the exchange rate of the country producing the input (or the product k). For now itis worth noting that the likely bias from violations of this assumption will be positive,leading us to over-state the rate of pass-through.
Figures 5 and 6 (for Ethiopia and Nigeria respectively) contain our estimates of thepass-through rate for all goods and all locations, with pass-through rates (within a des-tination location averaged across all products) again plotted nonparametrically againstlog source-to-destination distance (again in travel time units).11 It is important to stressthat the goal of Step 2 here is to estimate pass-through (ie ρk
od) separately for each des-tination d and product k. For expositional purposes only, Figures 5 and 6 plot averagesof these estimates, averaged across all products. A general tendency in these figures isfor the pass-through rate to be lower at destinations that are further distances from theproduct’s source. This is confirmed in column 1 of Tables 9 and 10 which show shows sig-nificant negative coefficients from the regression of pass-through estimates on log source-to-destination distance (again in travel time units). Another general tendency is for esti-mated pass-through to lie below one, often considerably below one; the average estimatedpass-through rate in our sample is approximately 0.5. The theory outlined above (indeedvirtually any oligopolstic model) places no restrictions on the pass-through rate exceptthat it be positive (a restriction that none of our estimates violate). But beyond this non-negativity restriction pass-through could be below or above one; our estimates suggestthat pass-through below one is a commonplace (and naturally our OLS estimates of thepass-through rate are likely to be, if anything, biased upwards).
While the primary goal of estimating pass-through rates is to feed into Step 3 of ouranalysis below, below we will also use our estimated pass-through rates (from Step 2here) to identify the competitiveness parameter φk
od prevailing in each location due to
the relationship between pass-through and competitiveness, ρkod =
(1 + δk
dφk
od
)−1
in our
model.
11For now we estimate only contemporaneous pass-through rates (though due to the high serial correla-tion in source prices these estimates are similar to those from lagged regressions). In future versions of thispaper we aim to estimated distributed lag specifications and hence trace out the entire impulse response inthe destination price of a change in the port price, that is both short-run and long-run pass-through.
25
4.3 Step 3: Using pass-through adjusted price gaps to measure how
distance affects the marginal costs of intranational trade.
In section 4.1, we detailed how the price gaps among trading pairs increased withthe (log) distance separating the trading pairs. However, this positive relationship is notdriven solely by the fact that the marginal costs of trading increase with distance. In addi-tion, intermediaries charge markups, and our model clarifies that the size of the markupmay be related to distance for two distinct reasons. The theoretical framework outlinedabove offers guidance here, and this is particularly easy to see in the case of constantdemand preferences (ie Assumption 4) for which, recall, Result 3 is
Pkdt − ρk
odtPkot
ρkodt
= τ(Xkodt) +
(1− ρkodt)
ρkodt
akdt. (22)
Recall that in Step 2 above we have obtained estimates of the time-constant (or averageover time) pass-through rate in each location d and product k, an estimate that we denoteby ρk
od. Using these estimates makes the estimation of τ(Xkodt) in equation (22) feasible.
Again, as in Step 2 above, an identification challenge is posed by the presence of theunobserved demand-shifter ak
dt on the right-hand side of this estimating equation. Wetherefore assume that ak
dt can be decomposed as follows: akdt = αk
t + αd + νkdt and, further,
that E[Xk
odtνkdt]= 0. This assumption requires that the variation in additive demand
shifters across destination locations (ie the variation, νkdt, that remains after macro-level
time-product effects, αkt , and destination effects, αd, are removed) is uncorrelated with
shifters to the marginal costs of trading across locations, Xkodt. Again, we require no re-
strictions at all on the multiplicative demand shifters, bkdt, from Equation (3) above.
With this assumption in place we can now state our main estimating equation foridentifying the extent to which distance affects the marginal costs of trading (ie how somemarginal cost shifter Xk
odt affects the marginal costs of trading, τ(Xkodt)):
Pkdt − ρk
odPkot
ρkod
= τ(Xkodt) + γk
t
1− ρkod
ρkod
+ γkt + γd
1− ρkod
ρkod
+ γd + εkdt, (23)
where ρkod is a consistent estimator of the pass-through rate ρk
od obtained in Step 2 above,
γkt is a product-time fixed-effect, γd is a destination fixed effect and εk
dt =(1−ρk
od)
ρkod
νkdt is an
error term for which E[Xk
odtνkdt]= 0.
26
The key attraction of this equation from an empirical perspective is that it describesa way in which price data across origin-destination pairs can be used, in conjunctionwith estimates ρk
od of the pass-through rate (obtained in Step 2 above), to estimate theimportance of shifters Xk
odt to the marginal costs of trading. Further, in principle the effectof Xk
odt on τ(Xkodt) can be estimated entirely non-parametrically. For example, a central
question in the study of trade costs concerns the extent to which distance increases themarginal costs of trading. Equation (23) implies that this relationship between distanceand the marginal costs of trade is revealed, despite the potential presence of market power
in the trading sector, by simply using ‘adjusted price gaps’ (ie Pkdt−ρk
odPkot
ρkod
) rather than price
gaps (ie Pkdt − Pk
ot) as the dependent variable.To gain intuition for this expression, consider the following. First, and dropping
the sub- and superscript notation for now without the risk of confusion, the size of themarkup, µ = (1− ρ)(a − τ − Po), is proportional to the gap between the choke price aand the total cost to the intermediary, τ + Po. For the case of incomplete pass through,ρ < 1, higher marginal costs of transportation raise prices and hence reduce the markupthat intermediaries choose to charge ( dµ
dXod|(m
θ )od= −(1− ρ) dτ
dXod). This channel implies
that the marginal costs of distance are understated in section 4.1 if ρ < 1 and overstated ifρ > 1. Second, the pass-through rate varies across space due to the competitiveness of the
distribution sector ( dµdXod|τ = −(a− τ − Po)
dρ
d(mθ )od
d(mθ )od
dXod). For the case of incomplete pass
through, if routes that reach interior locations far from major production/import loca-tions are less competitive, pass through rates will be smaller in the interior and markupswill be larger. This channel implies that the marginal costs of distance are overstated insection 4.1 if ρ < 1, but the direction of bias is ambiguous if ρ > 1.
Our aim in this section is to estimate the true marginal costs of distance by correctingfor these two biases using the adjusted price gap methodology described in equation(23). We obtain estimates of pass through rates, ρk
od, from section 4.2. Dividing the pricegap Pk
dt − Pkot by the pass-through rate purges the price gap of the first bias. Further
transforming the price gap by replacing Pkot with ρk
odPkot, as well as including two new sets
of independent variables, γkt
(1−ρk
od
ρkod
)and γd
(1−ρk
od
ρkod
), where γk
t and γd are product-time
and destination fixed effects, explicitly controls for the fact that markups may vary overspace due to different levels of competition.
Tables 5 and 6 (for Ethiopia and Nigeria respectively) present the results of these re-gressions, where we model the marginal costs of trading τ(Xk
odt) as a simple functionof (log) distance (from origin location o to destination location d), where distance is inunits of travel time as before. Column 1 reproduces the unadjusted price gap specifi-
27
cation shown in section 4.1 above (for the interpretation of the coefficients in this table,see below). Column 2 goes part-way towards adjusting these estimates for potentiallyvarying mark-ups over space by dividing the price gap through by the pass-throughrate. And column 3 estimates equation (23) in the manner suggested by our model—
that is, using the appropriate ‘adjusted price gap’ (ie Pkdt−ρk
odPkot
ρkod
) as the dependent variable
and controlling for estimates of pass-through adjusted demand shifters,γkt
(1−ρk
od
ρkod
)and
γd
(1−ρk
od
ρkod
).12 All specifications include product-time fixed effects. A consistent pattern
emerges across these three columns, in both Ethiopia and Nigeria, that is consistent withour model. First, the estimated coefficient on the (log) distance separating the tradingpairs at first rises (in column 2 relative to column 1) when only the adjusted price gap isused; but this coefficient then falls (in column 3 relative to column 2) when both correc-tions are applied. These results imply that the two biases discussed above are substantialin magnitude but, at least in the case of Nigeria and Ethiopia, cancel each other out tosome extent.
Columns 4-6 of Tables 5 and 6 allow for a more general specification of τ(Xkodt). We
allow the marginal costs of trading τ(Xkodt) to be functions of both the (log) distance be-
tween the source and destination, the log weight of a unit of the good in question andthe interactions of the two. In both Ethiopia and Nigeria, the marginal cots of distance issubstantially larger for heavier goods.
The estimated coefficients in column (3) of each table correspond, according to ourmethodology, to the estimated marginal costs of distance (in travel time units) along themean journey length in each country (approximately 6 and 8 hours respectively). Thesenumbers are substantially higher than the estimated total costs of distance. The estimatedcoefficients imply that the marginals costs of trade increase by 4.11 cents and 5.70 centsfor each additional unit of log-distance in Ethiopia and Nigeria respectively (comparedto 2.99 cents and 3.40 cents cents without the correction for spatial markup variation).
To interpret these estimates, consider the following. The least remote locations in oursample are approximately an hour of travel (ie 60 minutes or 4.1 log minutes) away fromthe source of production. (This travel time falls within the second percentile of the dis-tribution of route lengths in both samples). The most remote locations in our sample areapproximately twenty hours (ie 1200 minutes or 7.1 log minutes) of travel away from the
12As we are dividing by ρkod, results are very sensitive to estimated ρk
ods close to 1. Therefore, we win-sorize all the pass-through rates estimates that fall below 0.2. Our results are robust to this procedure andthe un-winsorized regressions are reported in tables 7 and 8.
28
source of production. (This travel time falls within the 99th percentile of route lengths inEthiopia and the 97th percentile in Nigeria). Therefore, the additional trade costs incurredby transporting goods to the most remote compared to the least remote locations (a differ-ence of 3 log minutes) is 12 cents in Ethiopia and 17 cents in Nigeria. The mean productobservation in our Ethiopia sample costs just over 40 cents. So the ad valorem equivalentof this relative cost of remoteness is 30 percent. The equivalent calculation for our Nige-ria sample (mean product cost of 1.17 dollars) suggests a relative cost of remoteness of 14percent. These are considerable costs of intranational trade.
To obtain a better sense of the magnitude, we can make rough comparisons to esti-mates of the marginal costs of distance from other sources. In a widely cited paper, Hum-mels (2001) uses freight cost data included in some countries customs records to estimatethe relationship between freight costs and distance for internationally traded goods. Ta-ble 3 of his paper reports estimates of the ad-valorem freight costs for imported goodsat relatively proximate and relatively far distances from the origin port. Taking simpledifferences of these ad valorem costs and dividing through by the change in log distanceprovides estimates comparable to the ad valorem numbers given above.13 The impliedincrease in ad valorem costs for an increase in distance of 3 log points are listed in thetable below for the seven countries for which Hummels’ could obtain data. The US cus-toms data is particularly rich, and allows separate estimates by mode of transport.
Implied ∆ad-valorem transport cost for ∆ lndistance of 3 units (percent)(by mode of transport for cargo of mean kg/$)
US (Truck from CAN) 2.0US (Rail from CAN) 2.7US (Ocean) 4.9US (Air) 14.6
Source: Authors’ calculation based on Table 3 in Hummels (2001).
The increase in ad valorem trade costs associated with an increase in distance of 3 logpoints was 30 percent in Ethiopia and 14 percent in Nigeria. Since almost all this tradetravels by road, the most easily comparable figures from Hummels are those for trucktraffic from Canadian provinces to the US border.14 The change in ad valorem trade costswith log distance in our two African samples is approximately 7 to 15 times larger thanthe change in ad valorem freight costs with log distance found for international tradeshipments from Canada to the US. Therefore, the Ethiopian and Nigerian numbers implyextremely large costs of intranational trade relative to international trade costs along the
13We take the difference between teh smallest and largest distance estimates reported by Hummels(2001). All of these estimates are for cargoes with the mean weight to value ratio.
14The import data record the province of origin in Canada and the district of entry into the US.
29
same mode of transport linking two developed countries. Interestingly, the estimates forNigeria are not too dissimilar to the estimates Hummels obtains for landlocked Paraguay,where imports must travel long distances over developing-country land routes prior toarriving at the dry port.
Figures 7 and 8 (for Ethiopia and Nigeria respectively) present our nonparametric esti-mates of the effect of distance (again, in travel time units) on the marginal costs of trading.By and large these plots confirm the parametric (linear) regression estimates referred toabove. Importantly, the ordering of slopes across three different methodologies (all pairs,trading pairs and markup-adjusted pairs) is the same nonparametrically (in these figures)as parametrically (in the regression coefficients reported above).
Tables 7 and 8 carry out a variety of important robustness checks. Columns 1-4 repeatthe estimation of the main specification shown in the previous table but with the inclu-sion of destination or destination-time fixed-effects. This does not substantially affect thecoefficient estimates. However, given the limited number of source locations, the destina-tion fixed effects are highly correlated with distance and hence our preferred specificationincludes only product-time fixed effects. Columns 5-6 control for destination-specific de-mand shocks in a more parsimonious manner by including controls for the log populationand log income per capita at the destination. Columns 7-8 do not winsorize the ρk
od esti-mates that fall below 0.2 as we do for the main specification to avoid dividing price gapsby numbers close to zero. Columns 9-10 use exchange rates deflated by the local inflationrate to instrument for origin prices in the estimation of ρk
od for the subsample of importgoods where bilateral deflated exchange rates explain origin price movements. Columns11-12 build on columns 9-10 but also include the deflated local currency oil price as anexplanatory variable in the pass through regression. The inclusion of the oil price ex-plicitly deals with the worry that shocks to oil prices (potentially due to exchange ratefluctuations) can alter both the origin prices and marginal costs of transport. Columns13-14 remove goods which show strong evidence of producer price setting behavior. Asall goods show price variation of space, we remove the four goods from the two samplesfor which nominal prices remain fixed for long periods of time (24 months or more). Fi-nally, Columns 15-16 remove price pairs where the destination location is less than 100minutes from the source location in case demand shocks are spatially correlated biasingour pass through estimates towards one for nearby locations.
30
4.4 Implications for the share surplus accruing to consumers, to inter-
mediaries, and to deadweight loss
Consider the following thought experiment. Due to tariff reductions or improvementsin international transportation and logistics, events often termed ‘globalization’, the portprice of a particular import falls by 20 percent. This price reduction creates an additionalamount of social surplus that will in part: (1) accrue as consumer surplus to consumerslocated at various points within the country, (2) accrue as profits to the intermediaries whoprovide consumers with the import goods, and (3) end up as deadweight loss associatedwith intermediaries using their market power to restrict supply. The model in section 3shows—in Result 4(c)—that under the assumption that demands are in the constant pass-through class, the pass through rate and the competitiveness index of any particular routeprovide sufficient statistics for estimating how the social surplus is distributed betweenconsumers, intermediaries and deadweight loss.
But where can estimates of these parameters be obtained?First, Result 2(b) above has already discussed a method for obtaining consistent esti-
mates of ρkd, namely equilibrium pass-through ρk
odt =
(1 + δk
dtφk
odt
)−1
under the additional
assumption (Assumption 5) that pass-through is constant over time within a product-destination market (because δk
dt and φkodt are constant).
Second, the formula
ρkod =
(1 +
δkd
φkod
)−1
(24)
suggests how the pass-through rate ρkod and the competitiveness index φk
od are connectedto one another and hence how estimates of ρk
od could be used to estimate φkod. Unfortu-
nately, in general there is no unique mapping between ρkod and φk
od. Indeed, as equation(24) above illustrates, in principle there are DT known values of ρk
od, but DT unknown val-ues of δk
d and another DT unknown values of φkod to be estimated. We therefore assume (in
Assumption 8) that the variation in the demand-side determinants of pass-through (ie theparameters δk
d) and the supply-side determinants of pass-through (ie the parameters φkod)
are sufficiently orthogonal over destination markets d and products k as to allow data onthe pass-through rate (ie an estimate of ρk
od) to identify φkod which is all that is required to
apply Result 4(b) and hence provide an answer to the question of how does the share oftotal surplus accruing to consumers vary with remoteness. However, as should be clear,the particular assumption made in Assumption 8 here is overly sufficient since it restrictsthere to be only D + T unknown parameters to be estimated from DT pass-through ρk
od
31
estimates.
Assumption 5. The demand parameter δkd is constant over destination locations d but can vary
freely across products k; that is, δkd=δk∀d. Similarly, the competitiveness index parameter φk
od isconstant over products k but is free to vary across destinations d; that is, φk
od=φd∀k.
This is a particularly stark assumption but one that is perhaps not an implausible first-pass. That is, because of possible economies of scale it seems plausible that the essentialvariation in the number of intermediaries and their competitive conduct (ie mk
od and θkd),
and hence the overall competiveness index φkod, across products and locations is primarily
across locations. We have in mind here a notion that large locations have many intermedi-aries supplying any given good. Likewise, while we allow the additive and multiplicativeshifters of demand (ie ak
dt and bkdt) to vary across locations, products and time, it seems
plausible that the second-order curvature parameter δkdt, the unique demand-side param-
eter that governs pass-through, is constant across locations and time. That said, becauseAssumption 8 is overly sufficient for identification, it is testable, and we will explore thisin future work.
We are finally ready to state the key result that describes an empirical procedure (andthe assumptions required for it to be accurate) to Question 4:
Result 4(c): Under Assumptions 1-8, a consistent estimator of the competitiveness index at adestination (ie φd), up to a scalar, can be obtained by estimating the following regression by OLS
Ξkod = γd + γk + γkζk
od + εkod, (25)
where Ξkod
.= ln( 1
ρkod
− 1) if ζkod = 1 and Ξk
od.= ln(1 − 1
ρkod
) if ζkod = 0, ρk
od is a consistent
estimator of the equilibrium pass-through rate obtained by applying Result 2(b), γd and γk aredestination- and product-specific fixed effects respectively, and εk
d is an error term. Normalizingsuch that the lowest value of φd = 1, a consistent estimator of φd is φd
.= eγd . Further, an
estimate of the ratio of intermediaries’ surplus ISkd(Q
kd) to consumer surplus CSk
d(Qkd) in the
market at destination location d for product k is given by
ISkd
CSkd
(Qk
d
)=
1
ρkod
+ e−γd − 1, (26)
and an estimate of the ratio of deadweight loss DWLkd(Q
kd) to consumer surplus CS(Qk
d) in this
32
market is given byDWLk
dt
CSkdt
(Qk
dt
)=
eγd + ρkod(e
γd − 1)
eγd
[eγd − ρk
od(eγd − 1)
] . (27)
Result 4(c) therefore describes a simple method for using estimated pass-through ratesρk
od, which we estimate separately by product k and destination location d, to infer how thedistibution of suprlus varies with remoteness. That is, despite the lack of access to con-sumption quantity data in this setting, we can use estimated pass-through rates, guidedby Result 4(c), to infer the share of total surplus—that is, the gains from trade—accruingto consumers, to intermediaries, and to deadweight loss in each market (ie product andlocation) in our sample.
Note also that, along the way to answering this question, Result 4(c) suggests a way inwhich the competitiveness index φd for each destination d can be identified. As outlinedin Result 4(c), the good and origin-destination specific pass through rates ρk
od providesufficient variation to estimate the unknown competitiveness indices for each destinationlocation d. Various methods can be used to obtain estimates of the competitiveness indexφd. Here, we follow Result 4(c) and transform the pass-through rates to linearize thisrelationship and then apply OLS to recover estimates of φd.
Figures 9 and 10 show non-parametric plots of how the competitiveness index varieswith (log) distance to the capital city, with higher values of the index representing greaterlevels of competition. If intermediaries compete a la Cournot, the competitiveness indexsimply corresponds to the number of middlemen serving a particular location. Column(2) of Tables 9 and 10 report descriptive regression of the competitiveness index at eachlocation against the log distance to the capital city (Addis Ababa or Lagos). The value ofthe competitiveness index clearly declines with distance form the capital in both Ethiopiaand Nigeria. More remote locations have a less competitive intermediary sector servingthem.
Result 4(c) above also described how estimates of the distribution of surplus followsimply from estimates of pass-through and competitiveness. Figures 11, 12 present non-parametric plots of the relationship between the ratio of intermediary profits to consumersurplus and distance for each good. Figures 13, 14 present similar plots for the otherratio of interest, the ratio of deadweight loss to consumer surplus. Finally, Figures 15, 16present similar plots for the share of consumer surplus in total suprlus. Columns (3) to (5)of Tables 9 and 10 report descriptive regressions of these ratios and shares against the logsource-to-destination distance. For both Ethiopia and Nigeria, the further a good must
33
travel to reach the consumer, the smaller share of the (partial equilibrium) surplus thataccrues to the consumer. The additional share of surplus going to consumers in the leastremote locations (1 hour away) compared to the most remote locations (20 hours away)is 7 percent in Ethiopia and 19 percent in Nigeria.
The normalization that the least competitive locations were served by monopolisttraders is not wholly innocuous (although necessary to avoid the dummy variable trapin our estimation strategy). Reassuringly, choosing other normalizations where the leastcompetitive location is more competitive than under monopoly changes the share goingto consumers but does not affect the slope of this share with respect to log distance (thekey comparison of interest).
The fact that consumers accrue only a fraction of the surplus generated by the abilityto purchase goods made in distant locations does not tell us how much surplus is createdin total. Our finding that the marginal costs of intranational trade are extremely highclearly implies that the total quantity of surplus will be smaller in more remote locations.In the extreme case, high marginal costs of trade may prevent goods from even reachingremote locations. In this scenario, a tariff cut at the border will not increase the socialsurplus of interior consumers at all.
Consumer price index enumerators record as missing a good that is unavailable ata particular location during a particular month. This information on product availabil-ity provides suggestive evidence that internal trade costs substantially reduce the totalquantity of social surplus generated by trade. If products cannot be found in locationsfar from the port or factory, neither consumers nor intermediaries in this location benefitfrom trade in this product. Figures 17 and 18 plot product availability (a binary variable)on the log source-to-destination distance for the subset of product-location pairs whereproduct is observed at least once in the sample. As expected, in both countries productavailability falls precipitously with distance from the factory or port.
5 Conclusion
This paper sets out to answer the question how large are intrnational trade costs in de-veloping countries? We find that the costs of distance appear to be under-estimated bystandard spatial price gap methods used to infer trade costs. The costs of distance ap-proximately double when we use discard uninformative price gaps, those price gaps forwhich neither of the pairs is a source location for the good in question. The costs of dis-tance approximately double again when spatial variation in mark-ups accounted for by
34
using a sufficient statistic (pass-through rates) to adjust price gaps�Our finding that the costs of intranational trade are extremely high (approximately
7 to 15 times larger than the freight costs for road transport between Canada and theUS), has obviously implications for consumer welfare. High intranational trade costsreduce the amount of potential surplus consumers can derive from purchasing goodsmade in distant locations. Of the surplus that remains once the costs of distance have beenaccounted for, it appears that a significant fraction does not actually accrue to consumers(and instead accrue to intermediaries and deadweight loss), and that this is especiallytrue in the most remote locations.
35
References
AHN, J., A. K. KHANDELWAL, AND S.-J. WEI (2011): “The Role of Intermediaries in Fa-cilitating Trade,” Journal of International Economics.
ALESSANDRIA, G., AND J. KABOSKI (2011): “Pricing-to-Market and the Failure of Abso-lute PPP,” American Economic Journal: Macroeconomics, 3(1), 91–127.
ANDERSON, J., AND E. VAN WINCOOP (2004): “Trade Costs,” Journal of Economic Litera-ture, 42(3), 691–751.
ANTRAS, P., AND A. COSTINOT (2011): “Intermediated Trade,” The Quarterly Journal ofEconomics, 126(3), 1319–1374.
ARKOLAKIS, C., A. COSTINOT, D. DONALDSON, AND A. RODRIGUEZ-CLARE (2012):“The Elusive Pro-Competitive Effects of Trade,” Discussion paper, unpublishedmanuscript.
ATKESON, A., AND A. BURSTEIN (2008): “Pricing-to-Market, Trade Costs, and Interna-tional Relative Prices,” The American Economic Review, 98(5), 1998–2031.
BARDHAN, P., D. MOOKHERJEE, AND M. TSUMAGARI (2011): “Middlemen Margins andGlobalization,” Discussion paper.
BERMAN, N., P. MARTIN, AND T. MAYER (2012): “How do different exporters react toexchange rate changes?,” The Quarterly Journal of Economics, 127(1), 437–492.
BRODA, C. M., AND D. E. WEINSTEIN (2008): “Understanding International Price Differ-ences Using Barcode Data,” Discussion paper.
BULOW, J. I., AND P. PFLEIDERER (1983): “A Note on the Effect of Cost Changes onPrices,” The Journal of Political Economy, 91(1), 182–185.
BURSTEIN, A., AND N. JAIMOVICH (2009): “Understanding Movements in Aggregate andProduct-Level Real Exchange Rates,” Discussion paper.
CHAU, N. H., H. GOTO, AND R. KANBUR (2009): “Middlemen, Non-Profits and Poverty,”Discussion paper.
DE LOECKER, J., P. K. GOLDBERG, A. K. KHANDELWAL, AND N. PAVCNIK (2012):“Prices, Markups and Trade Reform,” Working Paper 17925, National Bureau of Eco-nomic Research.
36
DONALDSON, D. (2011): “Railroads of the Raj: Estimating the Impact of TransportationInfrastructure,” Working Paper MIT.
EATON, J., AND S. KORTUM (2002): “Technology, Geography, and Trade,” Econometrica,70(5), 1741–1779.
EDMOND, C., V. MIDRIGAN, AND D. XU (2011): “Competition, Markups, and the Gainsfrom International Trade,” Discussion paper, unpublished manuscript.
ENGEL, C., AND J. H. ROGERS (1996): “How Wide is the Border?,” The American EconomicReview, 86(5), 1112–1125.
FACKLER, P. L., AND B. K. GOODWIN (2001): “Spatial Price Analysis,” Handbook of Agri-cultural Economics, 1, 971–1024.
FEENSTRA, R. (1989): “Symmetric pass-through of tariffs and exchange rates under im-perfect competition: an empirical test,” Journal of International Economics, 27(1-2), 25–45.
FEENSTRA, R. C., AND D. E. WEINSTEIN (2010): “Globalization, Markups, and the U.S.Price Level,” Working Paper 15749, National Bureau of Economic Research.
FITZGERALD, D., AND S. HALLER (2010): “Pricing-to-Market: Evidence from Plant-LevelPrices,” unpublished manuscript.
GOLDBERG, P. K., AND R. HELLERSTEIN (2008): “A Structural Approach to Explaining In-complete Exchange-Rate Pass-Through and Pricing-to-Market,” The American EconomicReview, 98(2), 423–429.
GOLDBERG, P. K., AND M. M. KNETTER (1997): “Goods Prices and Exchange Rates: WhatHave We Learned?,” Journal of Economic Literature, 35(3), 1243–1272.
HUMMELS, D. (2001): “Toward a Geography of Trade Costs,” Discussion paper, Mimeo,Purdue University.
KELLER, W., AND C. H. SHIUE (2007): “Markets in China and Europe on the Eve of theIndustrial Revolution,” The American Economic Review, pp. 1189–1216.
LI, N., G. GOPINATH, P. GOURINCHAS, AND C. HSIEH (2011): “International Prices,Costs and Markup Differences,” The American Economic Review, 101(6), 2450–2486.
MELITZ, M., AND G. OTTAVIANO (2008): “Market size, trade, and productivity,” Reviewof Economic studies, 75(1), 295–316.
37
NAKAMURA, E., AND D. ZEROM (2010): “Accounting for Incomplete Pass-Through,” Re-view of Economic Studies, 77(3), 1192–1230.
PAKES, A. (2008): “Theory and empirical work on imperfectly competitive markets,” Dis-cussion paper, National Bureau of Economic Research.
PARSLEY, D. C., AND S.-J. WEI (2001): “Explaining the Border Effect: The Role og Ex-change Rate Variability, Shipping Costs, and Geography,” Journal of International Eco-nomics, 55(1), 87.
SEADE, J. (1980): “On the Effects of Entry,” Econometrica: Journal of the Econometric Society,pp. 479–489.
SIMONOVSKA, I. (2010): “Income Differences and Prices of Tradables,” Discussion paper,UC Davis Working Paper.
SIMONOVSKA, I., AND M. WAUGH (2011a): “The Elasticity of Trade: Estimates and Evi-dence,” Discussion paper, National Bureau of Economic Research.
WEYL, E. G., AND M. FABINGER (2011): “A Restatement of the Theory of Monopoly,”Discussion paper.
38
Figures
Figure 1: Map of sample locations in Ethiopia
!
!!
!
!
!
!
!!
!
!
!
!
!
!
!!
!!
!!
! !!
!
!
!
!
!
!!
!
!
!
!
!
!!
!
!
!
!
!!
!
!
!
!
!
!
!
!
!
! !
!!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!!
!
!
!
!!
!
!
! !!
! !!
!
!!
!
!
!
#*
#*#*#*#*
#*
#*#*
#*
#*
#*
#*#*
#*#*
#*
#*
#*
#*#*#*#*#*
#*
#*
#* Source locations! Market observations
Primary roadsSecondary roads
#*
ss! Market observations
Note: Red triangles denote market locations at which prices are observed and blue circlesdenote origin locations (factories in the case of domestically produced goods, ports-of-entry in the case of foreign goods).
39
Figure 2: Map of sample locations in Nigeria
!
! !
!
!
!
!
!
!
!
! !!
!
!
!
!
!
!
!
!
!!
!
!
!
!
!
! !!
!
!
!
!
!
!!
#*#*
#*
#*#*#*#*#* #*#*#*#*#*#*#*#*#*#*#*#*#* #*#*
#*
#* Source locations! Market observations
Primary roadsSecondary roads
Note: Red triangles denote market locations at which prices are observed and blue circlesdenote origin locations (factories in the case of domestically produced goods, ports-of-entry in the case of foreign goods).
40
Figure 3: Intranational trade costs and distance: Ethiopia
0.0
5.1
Cos
ts o
f dis
tanc
e (2
001
US
$)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
all pairs price gaps trading pairs price gaps
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).All plots are semiparametric and include product−time fixed effects (Baltagi and Li, 2002).
Ethiopia: All Goods
Figure 4: Intranational trade costs and distance: Nigeria
0.0
5.1
Cos
ts o
f dis
tanc
e (2
001
US
$)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
all pairs price gaps trading pairs price gaps
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).All plots are semiparametric and include product−time fixed effects (Baltagi and Li, 2002).
Nigeria: All Goods
41
Figure 5: Estimated pass-through rates for all goods and distance: Ethiopia
.5.5
5.6
.65
.7.7
5P
ass−
thro
ugh
rate
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 6: Estimated pass-through rates for all goods and distance: Nigeria
0.2
.4.6
.8P
ass−
thro
ugh
rate
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
42
Figure 7: Marginal costs of distance: Ethiopia
0.1
.2C
ost o
f Dis
tanc
e (2
001
US
$)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
all pairs trading pairs µ−adjusted trading pairs
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).All plots are semiparametric and include product−time fixed effects. Adjusted gaps control for market power.
Ethiopia: All Goods
Figure 8: Marginal costs of distance: Nigeria
0.1
.2C
ost o
f Dis
tanc
e (2
001
US
$)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
all pairs trading pairs µ−adjusted trading pairs
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).All plots are semiparametric and include product−time fixed effects. Adjusted gaps control for market power.
Nigeria: All Goods
43
Figure 9: Competitiveness of intermediaries and distance: Ethiopia
1.8
22.
22.
42.
6
Rel
ativ
e co
mpe
titiv
enes
s in
dex
of in
term
edia
ries
at d
estin
atio
n m
arke
t
60 120 240 600 1200Distance from location to capital city (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 10: Competitiveness of intermediaries and distance: Nigeria
33.
54
4.5
5
Rel
ativ
e co
mpe
titiv
enes
s in
dex
of in
term
edia
ries
at d
estin
atio
n m
arke
t
60 120 240 600 1200Distance from location to capital city (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
44
Figure 11: Ratio of intermediary profit to consumer surplus and distance: Ethiopia
1.2
1.4
1.6
1.8
22.
2
Rel
ativ
e ra
tio o
f int
erm
edia
ry p
rofit
sto
con
sum
er s
urpl
us
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 12: Ratio of intermediary profit to consumer surplus and distance: Nigeria
11.
52
2.5
Rel
ativ
e ra
tio o
f int
erm
edia
ry p
rofit
sto
con
sum
er s
urpl
us
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
45
Figure 13: Ratio of deadweight loss to consumer surplus and distance: Ethiopia
.2.2
5.3
.35
.4
Rel
ativ
e ra
tio o
f dea
dwei
ght l
oss
to c
onsu
mer
sur
plus
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 14: Ratio of deadweight loss to consumer surplus and distance: Nigeria
.05
.1.1
5.2
Rel
ativ
e ra
tio o
f dea
dwei
ght l
oss
to c
onsu
mer
sur
plus
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
46
Figure 15: Share of consumer surplus in total surplus and distance: Ethiopia
.38
.4.4
2.4
4.4
6.4
8S
hare
of t
otal
sur
plus
acc
ruin
g to
con
sum
ers
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 16: Share of consumer surplus in total surplus and distance: Nigeria
.35
.4.4
5.5
.55
.6S
hare
of t
otal
sur
plus
acc
ruin
g to
con
sum
ers
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
47
Figure 17: Product availability: Ethiopia
.75
.8.8
5.9
.95
Pro
duct
Ava
ilabi
lity
(Bin
ary)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Ethiopia: All Goods
Figure 18: Product availability: Nigeria
.75
.8.8
5.9
.95
Pro
duct
Ava
ilabi
lity
(Bin
ary)
60 120 240 600 1200Distance from source location to destination market (minutes, log scale)
95% confidence intervals shown. Locally weighted polynomial (Epanechnikov kernel, bandwidth=0.5).
Nigeria: All Goods
48
Tables
49
Tabl
e1:
Tota
lint
rana
tion
altr
ade
cost
s,on
aver
age:
Ethi
opia
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Elec
tric
Bulb
Win
eD
rink
Det
erge
ntTo
iletS
oap
Dri
nkBa
tter
ies
-C
igar
ette
sPh
ilips
Sari
sM
eta
Beer
Zah
ira
Lux
Coc
aC
ola
Ever
eady
Rot
hman
s(4
0/60
w)
(750
cc)
(330
cc)
(50G
m)
(90G
m)
(300
cc)
Dry
cell
(pac
k)
Abs
olut
ePr
ice
Gap
0.03
40.
297
0.02
80.
006
0.01
80.
035
0.03
10.
113
(All
Pair
s)(0
.039
(0.2
80(0
.029
(0.0
05)
(0.0
19)
(0.0
35)
(0.0
48)
(0.2
38)
Pric
eG
ap0.
029
-0.0
330.
027
0.00
30.
005
0.05
80.
000
-0.0
19(T
radi
ngPa
irs)
(0.0
40)
(0.3
22)
(0.0
34)
(0.0
06)
(0.0
21)
(0.0
38)
(0.0
57)
(0.2
08)
Ori
gin
Pric
e0.
165
1.04
90.
283
0.05
50.
180
0.14
70.
171
0.89
2(0
.037
)(0
.107
)(0
.021
)(0
.004
)(0
.012
)(0
.019
)(0
.049
)(0
.121
)
Dis
tanc
eto
521.
9335
6.87
383.
1051
3.11
524.
0235
6.14
504.
0150
2.26
Sour
ce(m
inut
es)
(205
.85)
(177
.65)
(198
.66)
(199
.49)
(206
.85)
(197
.70)
(214
.74)
(208
.78)
Uni
tWei
ght(
Gm
)20
850
400
5090
350
4020
Obs
erva
tion
s44
5402
2434
0925
2238
1234
4246
4758
4630
4331
9126
2552
96
(9)
(10)
(11)
(12)
(13)
(14)
(15)
Wat
erBe
erH
air
Oil
Mot
orO
ilD
rink
Pen
Beer
Am
boH
arar
Zen
ith
Mob
ilPe
psi
Bic
Bede
le(5
00cc
)(3
30cc
)(3
30cc
)(1
Lt)
(300
cc)
(Bal
lPoi
nt)
(330
cc)
Abs
olut
ePr
ice
Gap
0.04
10.
030
0.06
70.
135
0.03
00.
011
0.03
0(A
llPa
irs)
(0.0
45)
(0.0
41)
(0.0
65)
(0.2
02)
(0.0
27)
(0.0
12)
(0.0
31)
Pric
eG
ap0.
070
0.03
80.
042
0.02
20.
036
0.00
40.
067
(Tra
ding
Pair
s)(0
.047
(0.0
39)
(0.0
79)
(0.1
94)
(0.0
33)
(0.0
14)
(0.0
31)
Ori
gin
Pric
e0.
133
0.26
30.
599
1.85
90.
152
0.09
50.
238
(0.0
17(0
.031
)(0
.060
)(0
.217
)(0
.023
)(0
.014
)(0
.025
)
Dis
tanc
eto
423.
1458
4.72
394.
8536
8.16
294.
4052
0.86
469.
55So
urce
(min
utes
)(1
83.1
6(2
40.3
3(1
84.1
5)(1
84.2
6)(1
57.0
9)(2
06.6
0)(2
14.6
0)
Uni
tWei
ght(
Gm
)10
0040
035
012
0032
020
400
Obs
erva
tion
s41
8829
2465
9142
9193
2216
7148
4802
4702
6624
6239
Not
es:
Row
1us
esda
tafr
omal
lloc
atio
npa
irs.
Row
3on
lyus
esda
tafr
om“t
radi
ngpa
irs”
,eg
pair
sw
here
one
ofth
elo
cati
ons
isei
ther
the
fact
ory
loca
tion
for
that
prod
ucto
rth
epo
rtof
entr
y.Pr
ices
are
defla
ted
byth
eav
erag
eof
the
prop
orti
onal
pric
ech
ange
for
each
good
atit
sor
igin
loca
tion
.R
ealp
rice
sar
eco
nver
ted
into
US
Dol
lars
usin
gth
epr
evai
ling
exch
ange
rate
duri
ngth
eba
sepe
riod
(Jan
uary
2001
).St
anda
rder
rors
inpa
rent
hese
s.
50
Tabl
e2:
Tota
lint
rana
tion
altr
ade
cost
s,on
aver
age:
Nig
eria
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Dri
nkC
emen
tBa
byFo
odD
eter
gent
Tea
Mar
gari
neSa
rdin
esBe
erC
emen
tBo
urnv
ita
Elep
hant
Cer
elac
Om
oLi
pton
Blue
Band
Titu
sG
uine
ssN
iger
cem
Abs
olut
ePr
ice
Gap
0.10
60.
606
0.14
30.
054
0.04
10.
124
0.04
10.
072
0.71
1(A
llPa
irs)
(0.1
28)
(0.5
53)
(0.1
42)
(0.0
77)
(0.1
31)
(0.1
18)
(0.0
38)
(0.1
08)
(1.0
23)
Pric
eG
ap0.
039
0.49
00.
039
-0.0
210.
000
0.04
6-0
.007
0.03
10.
360
(Tra
ding
Pair
s)(0
.119
)(0
.672
)(0
.161
)(0
.073
)(0
.037
)(0
.155
)(0
.047
)(0
.103
)(0
.932
)
Ori
gin
Pric
e1.
231
4.53
41.
259
0.35
80.
272
0.83
00.
366
0.42
43.
983
(0.0
48)
(0.8
55)
(0.1
34)
(0.0
71)
(0.0
21)
(0.1
77)
(0.0
75)
(0.0
40)
(0.2
38)
Dis
tanc
eto
600.
9356
6.98
579.
4257
2.69
643.
6558
4.71
585.
3835
6.06
458.
85So
urce
(min
utes
)(2
89.2
7)(2
77.8
1)(2
79.1
2)(2
68.1
7)(2
89.9
3)(2
74.2
3)(2
71.9
2)(2
76.5
6)(2
72.9
5)
Uni
tWei
ght(
Gm
)45
050
000
400
100
226
250
125
300
5000
0
Obs
erva
tion
s16
852
4656
348
191
3066
812
182
5338
453
498
5876
1071
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
Dri
nkD
rink
Dri
nkSu
gar
Pow
der.
Milk
Evap
.Milk
Evap
.Milk
Cig
aret
tes
Cok
eBt
l.C
oke
Can
Milo
StLo
uis
Peak
Peak
Car
nati
onB&
H
Abs
olut
ePr
ice
Gap
0.00
90.
021
0.10
50.
048
0.09
10.
021
0.04
00.
039
(All
Pair
s)(0
.008
)(0
.030
)(0
.093
)(0
.052
)(0
.088
)(0
.020
)(0
.101
)(0
.037
)
Pric
eG
ap-0
.002
-0.0
020.
007
0.01
00.
010
-0.0
22-0
.003
0.02
8(T
radi
ngPa
irs)
(0.0
13)
(0.0
27)
(0.1
25)
(0.0
56)
(0.1
10)
(0.0
29)
(0.0
58)
(0.0
39)
Ori
gin
Pric
e0.
128
0.25
41.
375
0.38
61.
368
0.32
00.
206
0.44
0(0
.012
)(0
.012
)(0
.059
)(0
.044
)(0
.081
)(0
.064
)(0
.011
)(0
.019
)
Dis
tanc
eto
623.
5962
8.91
623.
5958
5.15
613.
3157
9.17
654.
6554
8.18
Sour
ce(m
inut
es)
(292
.96)
(297
.96)
(291
.06)
(291
.06)
(293
.37)
(272
.61)
(273
.41)
(283
.16)
Uni
tWei
ght(
Gm
)30
020
045
090
040
017
017
030
0
Obs
erva
tion
s83
4871
7880
5323
260
6946
5383
044
5281
94
Not
es:
Row
1us
esda
tafr
omal
llo
cati
onpa
irs.
Row
3on
lyus
esda
tafr
om“t
radi
ngpa
irs”
,eg
pair
sw
here
one
ofth
elo
cati
ons
isei
ther
the
fact
ory
loca
tion
for
that
prod
ucto
rth
epo
rtof
entr
y.Pr
ices
are
defla
ted
byth
eav
erag
eof
the
prop
orti
onal
pric
ech
ange
for
each
good
atit
sor
igin
loca
tion
.R
ealp
rice
sar
eco
nver
ted
into
US
Dol
lars
usin
gth
epr
evai
ling
exch
ange
rate
duri
ngth
eba
sepe
riod
(Jan
uary
2001
).St
anda
rder
rors
inpa
rent
hese
s.
51
Tabl
e3:
Esti
mat
ing
the
Tota
lEff
ecto
fDis
tanc
e:Et
hiop
ia
(1)
(2)
Abs
olut
ePr
ice
Gap
Pric
eG
ap(A
llPa
irs)
(Tra
ding
Pair
s)
Log
dist
ance
betw
een
0.01
30**
*0.
0299
***
loca
tion
s(m
inut
es)
(0.0
0049
1)(0
.001
54)
Obs
erva
tion
s4,
302,
532
100,
905
R-s
quar
ed0.
356
0.25
8
Not
es:
All
colu
mns
regr
ess
log
dist
ance
ona
mea
sure
ofth
epr
ice
gap
for
each
good
and
pair
oflo
cati
ons.
Col
umns
1us
esda
taon
the
abso
lute
pric
ega
pbe
twee
nal
lloc
atio
npa
irs.
Col
umn
2us
esda
taon
the
actu
alpr
ice
gap
be-
twee
n“t
radi
ngpa
irs”
,eg
dest
inat
ion
pric
em
inus
orig
inpr
ice
for
pair
sw
here
one
ofth
elo
cati
ons
isei
ther
the
fact
ory
loca
tion
for
that
prod
uct
orth
epo
rtof
entr
y.Pr
ices
are
defla
ted
byth
eav
erag
eof
the
prop
orti
onal
pric
ech
ange
for
each
good
atit
sor
igin
loca
tion
.Rea
lpri
ces
are
conv
erte
din
toU
SD
olla
rsus
ing
the
prev
ailin
gex
chan
gera
tedu
ring
the
base
peri
od(J
anua
ry20
01).
All
regr
essi
ons
incl
ude
tim
e-pr
oduc
tfix
edef
fect
s.Ti
me-
prod
uct
clus
tere
dst
an-
dard
erro
rsin
pare
nthe
ses.
*si
gnifi
cant
at10
perc
entl
evel
,**
at5
perc
enta
nd**
*at
1pe
rcen
t.
52
Tabl
e4:
Esti
mat
ing
the
Tota
lEff
ecto
fDis
tanc
e:N
iger
ia
(1)
(2)
Abs
olut
ePr
ice
Gap
Pric
eG
ap(A
llPa
irs)
(Tra
ding
Pair
s)
Log
dist
ance
betw
een
0.02
37**
*0.
0340
***
loca
tion
s(m
inut
es)
(0.0
0259
)(0
.005
07)
Obs
erva
tion
s38
5,13
623
,520
R-s
quar
ed0.
490
0.48
9
Not
es:
All
colu
mns
regr
ess
log
dist
ance
ona
mea
sure
ofth
epr
ice
gap
for
each
good
and
pair
oflo
cati
ons.
Col
umns
1us
esda
taon
the
abso
lute
pric
ega
pbe
twee
nal
lloc
atio
npa
irs.
Col
umn
2us
esda
taon
the
actu
alpr
ice
gap
be-
twee
n“t
radi
ngpa
irs”
,eg
dest
inat
ion
pric
em
inus
orig
inpr
ice
forp
airs
whe
reon
eof
the
loca
tion
sis
eith
erth
efa
ctor
ylo
cati
onfo
rth
atpr
oduc
tor
the
port
ofen
try.
Pric
esar
ede
flate
dby
the
aver
age
ofth
epr
opor
tion
alpr
ice
chan
gefo
rea
chgo
odat
its
orig
inlo
cati
on.R
ealp
rice
sar
eco
nver
ted
into
US
Dol
lars
usin
gth
epr
evai
ling
exch
ange
rate
duri
ngth
eba
sepe
riod
(Jan
uary
2001
).A
llre
gres
sion
sin
clud
eti
me-
prod
uct
fixed
effe
cts.
Tim
e-pr
oduc
tcl
uste
red
stan
-da
rder
rors
inpa
rent
hese
s.*s
igni
fican
tat1
0pe
rcen
tlev
el,*
*at5
perc
enta
nd**
*at
1pe
rcen
t.
53
Tabl
e5:
Esti
mat
ing
the
Prim
itiv
eEf
fect
ofD
ista
nce:
Ethi
opia
(1)
(2)
(3)
(4)
(5)
(6)
Pric
eG
apPr
ice
Gap
/ρ
k odA
djus
ted
Pric
eG
apPr
ice
Gap
Pric
eG
ap/
ρk od
Adj
uste
dPr
ice
Gap
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
Log
dist
ance
to0.
0289
***
0.05
48**
*0.
0411
***
-0.0
459*
**-0
.065
6***
-0.0
906*
**so
urce
(min
utes
)(0
.001
47)
(0.0
0256
)(0
.002
46)
(0.0
0430
)(0
.008
50)
(0.0
0949
)
Log
dist
ance×
0.01
51**
*0.
0242
***
0.02
50**
*Lo
gw
eigh
t(0
.001
09)
(0.0
0193
)(0
.002
13)
Tim
e-Pr
oduc
tFE
Yes
Yes
Yes
Yes
Yes
Yes
Tim
e-Pr
oduc
t×1−
ρk od
ρk od
No
No
Yes
No
No
Yes
Des
tina
tion×
1−ρ
k od
ρk od
No
No
Yes
No
No
Yes
Obs
erva
tion
s10
0762
1007
6210
0762
1007
6210
0762
1007
62R
-squ
ared
0.25
80.
283
0.93
30.
272
0.28
80.
933
Not
es:C
olum
n1
regr
esse
slo
gdi
stan
ceon
the
pric
ega
pbe
twee
ntr
adin
gpa
irs
asin
tabl
e3.
Col
umn
2re
gres
ses
log
dist
ance
onth
epr
ice
gap
divi
ded
byth
ees
tim
ated
pass
thro
ugh
rate
sfr
omse
ctio
n4.
2,P
k ds−
Pk os
ρk od
,in
orde
rto
acco
untf
orth
ech
ange
inm
arku
psin
duce
dby
tran
spor
tcos
ts.C
olum
n3
uses
the
tran
sfor
med
pric
ega
pP
k ds−
ρk od
Pk os
ρk od
and
addi
tion
ally
incl
udes
tim
e-pr
oduc
tan
dde
stin
atio
nfix
edef
fect
sm
ulti
plie
dby
1−ρ
k od
ρk od
inor
der
toco
ntro
lfo
r
omit
ted
vari
able
bias
due
toth
ele
velo
fmar
ketp
ower
cova
ryin
gw
ith
dist
ance
.Col
umns
4,5
and
6re
peat
the
anal
ysis
buta
lso
incl
udin
glo
gdi
stan
cein
tera
cted
wit
hlo
gw
eigh
tfor
each
good
.All
regr
essi
ons
incl
ude
tim
e-pr
oduc
tfixe
def
fect
s.Ti
me-
prod
uctc
lust
ered
stan
dard
erro
rsin
pare
nthe
ses.
*si
gnifi
cant
at10
perc
entl
evel
,**
at5
perc
enta
nd**
*at
1pe
rcen
t.
54
Tabl
e6:
Esti
mat
ing
the
Prim
itiv
eEf
fect
ofD
ista
nce:
Nig
eria
(1)
(2)
(3)
(4)
(5)
(6)
Pric
eG
apPr
ice
Gap
/ρ
k odA
djus
ted
Pric
eG
apPr
ice
Gap
Pric
eG
ap/
ρk od
Adj
uste
dPr
ice
Gap
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
(Tra
ding
Pair
s)(T
radi
ngPa
irs)
Log
dist
ance
to0.
0343
***
0.06
37**
*0.
0570
***
-0.2
71**
*-0
.541
***
-0.3
84**
*so
urce
(min
utes
)(0
.005
29)
(0.0
102)
(0.0
0862
)(0
.022
3)(0
.040
2)(0
.032
1)
Log
dist
ance×
0.04
85**
*0.
0960
***
0.06
62**
*Lo
gw
eigh
t(0
.004
07)
(0.0
0728
)(0
.005
61)
Tim
e-Pr
oduc
tFE
Yes
Yes
Yes
Yes
Yes
Yes
Tim
e-Pr
oduc
t×1−
ρk od
ρk od
No
No
Yes
No
No
Yes
Des
tina
tion×
1−ρ
k od
ρk od
No
No
Yes
No
No
Yes
Obs
erva
tion
s23
084
2308
423
084
2308
423
084
2308
4R
-squ
ared
0.50
40.
399
0.96
40.
544
0.44
50.
965
Not
es:C
olum
n1
regr
esse
slo
gdi
stan
ceon
the
pric
ega
pbe
twee
ntr
adin
gpa
irs
asin
tabl
e3.
Col
umn
2re
gres
ses
log
dist
ance
onth
epr
ice
gap
divi
ded
byth
ees
tim
ated
pass
thro
ugh
rate
sfr
omse
ctio
n4.
2,P
k ds−
Pk os
ρk od
,in
orde
rto
acco
untf
orth
ech
ange
inm
arku
psin
duce
dby
tran
spor
tcos
ts.C
olum
n3
uses
the
tran
sfor
med
pric
ega
pP
k ds−
ρk od
Pk os
ρk od
and
addi
tion
ally
incl
udes
tim
e-pr
oduc
tfix
edef
fect
sm
ulti
plie
dby
1−ρ
k od
ρk od
inor
der
toco
ntro
lfor
omit
ted
vari
able
bias
due
toth
ele
velo
fmar
ketp
ower
cova
ryin
gw
ith
dist
ance
.C
olum
ns4,
5an
d6
repe
atth
ean
alys
isbu
tals
oin
clud
ing
log
dist
ance
inte
ract
edw
ith
log
wei
ghtf
orea
chgo
od.A
llre
gres
sion
sin
clud
eti
me-
prod
uctfi
xed
effe
cts.
Tim
e-pr
oduc
tclu
ster
edst
anda
rder
rors
inpa
rent
hese
s.*
sign
ifica
ntat
10pe
rcen
tlev
el,*
*at
5pe
rcen
tand
***
at1
perc
ent.
55
Tabl
e7:
Esti
mat
ing
the
Prim
itiv
eEf
fect
ofD
ista
nce:
Rob
ustn
ess
Che
cks
Ethi
opia
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Des
tina
tion
Des
tina
tion
-Yea
rC
ontr
ols
For
Cit
yN
otW
inso
rizi
ngFi
xed
Effe
cts
Fixe
dEf
fect
sSi
zean
dIn
com
ePa
ssTh
roug
hR
ates
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Log
dist
ance
to0.
0175
***
0.02
91**
*0.
0176
***
0.03
060.
0289
***
0.03
79so
urce
(min
utes
)(0
.001
31)
(0.0
0180
)(0
.001
30)
(.)(0
.001
47)
(.)
Tim
e-Pr
od.F
EYe
sYe
sYe
sYe
sYe
sYe
sYe
sYe
s
Tim
e-Pr
od.×
1−ρ
k od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
tina
tion
FEYe
sYe
sN
oN
oN
oN
oN
oN
o
Des
t.×1−
ρk od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
t.-Ye
arFE
No
No
Yes
Yes
No
No
No
No
Des
t.-Ye
ar×
1−ρ
k od
ρk od
No
No
No
Yes
No
No
No
No
Obs
erva
tion
s10
0762
1007
6210
0762
1007
6210
0762
1007
62R
-squ
ared
0.29
40.
934
0.31
80.
941
0.25
80.
995
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
Usi
ngEx
chan
geR
ates
Con
trol
sFo
rO
ilR
emov
ing
Goo
dsw
ith
Rem
ovin
gLo
cati
ons
asIV
’sin
ρk od
Pric
ein
ρk od
Prod
ucer
Pric
eSe
ttin
g<1
00M
inut
esA
way
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Log
dist
ance
to0.
0289
***
0.03
98**
*0.
0289
***
0.04
08**
*0.
0298
***
0.04
03**
*0.
0423
***
0.05
70**
*so
urce
(min
utes
)(0
.001
47)
(0.0
0240
)(0
.001
47)
(0.0
0256
)(0
.001
77)
(0.0
0293
)(0
.002
29)
(0.0
0367
)
Tim
e-Pr
od.F
EYe
sYe
sYe
sYe
sYe
sYe
sYe
sYe
s
Tim
e-Pr
od.×
1−ρ
k od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
t.×1−
ρk od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Obs
erva
tion
s10
0762
1007
6210
0762
1007
6279
218
7921
896
059
9605
9R
-squ
ared
0.25
80.
934
0.25
80.
933
0.24
60.
933
0.25
40.
933
Not
es:
For
each
spec
ifica
tion
pair,
the
first
colu
mn
regr
esse
slo
gdi
stan
ceon
the
pric
ega
pbe
twee
ntr
adin
gpa
irs
and
the
seco
ndco
lum
n
regr
esse
sth
etr
ansf
orm
edpr
ice
gap(P
k ds−
ρk od
Pk os)/
ρk od
onlo
gdi
stan
cean
dad
diti
onal
lyin
clud
esti
me-
prod
ucta
ndde
stin
atio
nfix
edef
fect
s
mul
tipl
ied
by(1−
ρk od)/
ρk od
inor
der
toco
ntro
lfo
rom
itte
dva
riab
lebi
asdu
eto
the
leve
lof
mar
ket
pow
erco
vary
ing
wit
hdi
stan
ce.
All
regr
essi
ons
incl
ude
tim
e-pr
oduc
tfix
edef
fect
s.C
olum
ns1
to2
incl
ude
dest
inat
ion
fixed
effe
cts.
Col
umns
3-4
repl
ace
the
dest
inat
ion
fixed
effe
cts
and
inte
ract
ions
wit
hde
stin
atio
n-ye
arfix
edef
fect
san
din
tera
ctio
ns.
Col
umns
5-6
incl
ude
cont
rols
for
the
dest
inat
ion
loca
tion
log
popu
lati
onan
dlo
gin
com
epe
rca
pita
.C
olum
ns7-
8ut
ilize
raw
ρk od
esti
mat
esth
atha
veno
tbe
enw
inso
rize
dbe
low
0.2.
Col
umns
9-10
use
defla
ted
exch
ange
rate
sto
inst
rum
ent
for
orig
inpr
ices
inth
ees
tim
atio
nof
ρk od
for
the
subs
ampl
eof
impo
rtgo
ods
whe
rebi
late
rald
eflat
edex
chan
gera
tes
expl
ain
orig
inpr
ice
mov
emen
ts.C
olum
ns11
-12
build
onco
lum
ns9-
10bu
tals
oin
clud
eth
ede
flate
dlo
calc
urre
ncy
oilp
rice
asan
expl
anat
ory
vari
able
inth
epa
ssth
roug
hre
gres
sion
.Col
umns
13-1
4re
mov
ego
ods
whi
chsh
owst
rong
evid
ence
ofpr
oduc
erpr
ice
sett
ing
beha
vior
.Col
umns
15-1
6re
mov
epr
ice
pair
sw
here
the
dest
inat
ion
loca
tion
isle
ssth
an10
0m
inut
esfr
omth
eso
urce
loca
tion
.Tim
e-pr
oduc
tcl
uste
red
stan
dard
erro
rsin
pare
nthe
ses.
*si
gnifi
cant
at10
perc
entl
evel
,**
at5
perc
enta
nd**
*at
1pe
rcen
t.
56
Tabl
e8:
Esti
mat
ing
the
Prim
itiv
eEf
fect
ofD
ista
nce:
Rob
ustn
ess
Che
cks
Nig
eria
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Des
tina
tion
Des
tina
tion
-Tim
eC
ontr
ols
For
Cit
yN
otW
inso
rizi
ngFi
xed
Effe
cts
Fixe
dEf
fect
sSi
zean
dIn
com
ePa
ssTh
roug
hR
ates
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Log
dist
ance
to0.
0858
***
0.07
54*
0.03
31*
0.05
410.
0343
***
0.05
29**
*so
urce
(min
utes
)(0
.018
0)(0
.038
6)(0
.018
2)(0
.041
2)(0
.005
29)
(0.0
0711
)
Tim
e-Pr
od.F
EYe
sYe
sYe
sYe
sYe
sYe
sYe
sYe
s
Tim
e-Pr
od.×
1−ρ
k od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
tina
tion
FEYe
sYe
sN
oN
oN
oN
oN
oN
o
Des
t.×1−
ρk od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
t.-Ti
me
FEN
oN
oYe
sYe
sN
oN
oN
oN
o
Des
t.-Ti
me×
1−ρ
k od
ρk od
No
No
No
Yes
No
No
No
No
Obs
erva
tion
s23
084
2308
423
520
2308
423
084
2308
4R
-squ
ared
0.51
80.
964
0.57
90.
975
0.50
40.
999
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
Usi
ngEx
chan
geR
ates
Con
trol
sFo
rO
ilR
emov
ing
Goo
dsw
ith
Rem
ovin
gLo
cati
ons
asIV
’sin
ρk od
Pric
ein
ρk od
Prod
ucer
Pric
eSe
ttin
g<1
00M
inut
esA
way
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Pric
eG
apA
dj.G
apPr
ice
Gap
Adj
.Gap
Log
dist
ance
to0.
0343
***
0.06
26**
*0.
0351
***
0.05
89**
*0.
0274
***
0.04
35**
*so
urce
(min
utes
)(0
.005
29)
(0.0
0929
)(0
.005
41)
(0.0
0884
)(0
.005
08)
(0.0
0895
)
Tim
e-Pr
od.F
EYe
sYe
sYe
sYe
sYe
sYe
sYe
sYe
s
Tim
e-Pr
od.×
1−ρ
k od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Des
t.×1−
ρk od
ρk od
No
Yes
No
Yes
No
Yes
No
Yes
Obs
erva
tion
s23
084
2308
422
676
2267
622
212
2221
2R
-squ
ared
0.50
40.
967
0.50
40.
964
0.52
30.
964
Not
es:
For
each
spec
ifica
tion
pair,
the
first
colu
mn
regr
esse
slo
gdi
stan
ceon
the
pric
ega
pbe
twee
ntr
adin
gpa
irs
and
the
seco
ndco
lum
n
regr
esse
sth
etr
ansf
orm
edpr
ice
gap(P
k ds−
ρk od
Pk os)/
ρk od
onlo
gdi
stan
cean
dad
diti
onal
lyin
clud
esti
me-
prod
ucta
ndde
stin
atio
nfix
edef
fect
s
mul
tipl
ied
by(1−
ρk od)/
ρk od
inor
der
toco
ntro
lfo
rom
itte
dva
riab
lebi
asdu
eto
the
leve
lof
mar
ket
pow
erco
vary
ing
wit
hdi
stan
ce.
All
regr
essi
ons
incl
ude
tim
e-pr
oduc
tfix
edef
fect
s.C
olum
ns1
to2
incl
ude
dest
inat
ion
fixed
effe
cts.
Col
umns
3-4
repl
ace
the
dest
inat
ion
fixed
effe
cts
and
inte
ract
ions
wit
hde
stin
atio
n-ti
me
fixed
effe
cts
and
inte
ract
ions
.C
olum
ns5-
6in
clud
eco
ntro
lsfo
rth
ede
stin
atio
nlo
cati
onlo
g
popu
lati
onan
dlo
gin
com
epe
rca
pita
.C
olum
ns7-
8ut
ilize
raw
ρk od
esti
mat
esth
atha
veno
tbe
enw
inso
rize
dbe
low
0.2.
Col
umns
9-10
use
defla
ted
exch
ange
rate
sto
inst
rum
ent
for
orig
inpr
ices
inth
ees
tim
atio
nof
ρk od
for
the
subs
ampl
eof
impo
rtgo
ods
whe
rebi
late
rald
eflat
edex
chan
gera
tes
expl
ain
orig
inpr
ice
mov
emen
ts.C
olum
ns11
-12
build
onco
lum
ns9-
10bu
tals
oin
clud
eth
ede
flate
dlo
calc
urre
ncy
oilp
rice
asan
expl
anat
ory
vari
able
inth
epa
ssth
roug
hre
gres
sion
.Col
umns
13-1
4re
mov
ego
ods
whi
chsh
owst
rong
evid
ence
ofpr
oduc
erpr
ice
sett
ing
beha
vior
.Col
umns
15-1
6re
mov
epr
ice
pair
sw
here
the
dest
inat
ion
loca
tion
isle
ssth
an10
0m
inut
esfr
omth
eso
urce
loca
tion
.Tim
e-pr
oduc
tcl
uste
red
stan
dard
erro
rsin
pare
nthe
ses.
*si
gnifi
cant
at10
perc
entl
evel
,**
at5
perc
enta
nd**
*at
1pe
rcen
t.
57
Tabl
e9:
Reg
ress
ing
Pass
-Thr
ough
Rat
es,C
ompe
titi
vene
ssan
dSu
rplu
sM
easu
res
onD
ista
nce:
Ethi
opia
Pass
-Thr
ough
Com
peti
tive
ness
Rat
ioof
Rat
ioof
Con
sum
er’s
Shar
eR
ates
Inde
xof
Inte
rmed
iary
Profi
tsD
eadw
eigh
tLos
sof
Tota
lSur
plus
Inte
rmed
iari
esto
Con
sum
erSu
rplu
sto
Con
sum
erSu
rplu
s(G
ood-
Loca
tion
)(A
llLo
cati
ons)
(Goo
d-Lo
cati
on)
(Goo
d-Lo
cati
on)
(Goo
d-Lo
cati
on)
Log
dist
ance
betw
een
-0.0
557*
**0.
284*
**0.
0283
***
-0.0
244*
**so
urce
and
dest
inat
ion
(0.0
155)
(0.0
606)
(0.0
0888
)(0
.007
69)
Log
dist
ance
betw
een
-0.2
30**
loca
tion
and
capi
tal
(0.1
06)
Con
stan
t0.
907*
**3.
459*
**0.
161
0.15
6***
0.55
3***
(0.0
913)
(0.6
21)
(0.3
59)
(0.0
533)
(0.0
460)
Obs
erva
tion
s14
1810
014
1814
1814
18R
-squ
ared
0.00
70.
027
0.01
40.
007
0.00
6
Not
es:
Col
umns
regr
ess
log
dist
ance
ones
tim
ates
ofpa
ss-t
hrou
ghra
tes,
com
peti
tive
ness
,the
rati
oof
inte
rmed
iary
profi
tto
cons
umer
surp
lus,
the
rati
oof
dead
wei
ghtl
oss
toco
nsum
ersu
rplu
san
dth
esh
are
ofco
nsum
ersu
rplu
sin
tota
lsur
plus
.Sta
ndar
der
rors
inpa
rent
hese
s.*
sign
ifica
ntat
10pe
rcen
tlev
el,*
*at
5pe
rcen
tand
***
at1
perc
ent.
58
Tabl
e10
:Reg
ress
ing
Pass
-Thr
ough
Rat
es,C
ompe
titi
vene
ssan
dSu
rplu
sM
easu
res
onD
ista
nce:
Nig
eria
Pass
-Thr
ough
Com
peti
tive
ness
Rat
ioof
Rat
ioof
Con
sum
er’s
Shar
eR
ates
Inde
xof
Inte
rmed
iary
Profi
tsD
eadw
eigh
tLos
sof
Tota
lSur
plus
Inte
rmed
iari
esto
Con
sum
erSu
rplu
sto
Con
sum
erSu
rplu
s(G
ood-
Loca
tion
)(A
llLo
cati
ons)
(Goo
d-Lo
cati
on)
(Goo
d-Lo
cati
on)
(Goo
d-Lo
cati
on)
Log
dist
ance
betw
een
-0.1
06*
0.33
6***
0.05
56**
*-0
.065
7***
sour
cean
dde
stin
atio
n-0
.054
4-0
.116
-0.0
0654
-0.0
186
Log
dist
ance
betw
een
-0.7
07**
*lo
cati
onan
dca
pita
l-0
.169
Con
stan
t1.
065*
**8.
004*
**-0
.010
3-0
.188
***
0.84
1***
-0.3
39-1
.012
-0.7
3-0
.038
8-0
.119
Obs
erva
tion
s48
936
489
489
489
R-s
quar
ed0.
011
0.15
0.01
90.
085
0.02
3
Not
es:
Col
umns
regr
ess
log
dist
ance
ones
tim
ates
ofpa
ss-t
hrou
ghra
tes,
com
peti
tive
ness
,the
rati
oof
inte
rmed
iary
profi
tto
cons
umer
surp
lus,
the
rati
oof
dead
wei
ghtl
oss
toco
nsum
ersu
rplu
san
dth
esh
are
ofco
nsum
ersu
rplu
sin
tota
lsur
plus
.Sta
ndar
der
rors
inpa
rent
hese
s.*
sign
ifica
ntat
10pe
rcen
tlev
el,*
*at
5pe
rcen
tand
***
at1
perc
ent.
59
Tabl
e11
:Reg
ress
ing
Prod
uctA
vaila
bilit
yon
Dis
tanc
e:Et
hiop
ia
(1)
(2)
(3)
Prod
uctA
vaila
bilit
yat
Leve
lofT
ime-
Prod
uct-
Loca
tion
(1=P
rice
Rec
orde
d,2=
No
Pric
eR
ecor
ded)
Log
dist
ance
to-0
.097
0***
-0.0
941*
**-0
.100
***
sour
ce(m
inut
es)
(0.0
0287
)(0
.002
56)
(0.0
0409
)
Con
stan
t1.
404*
**1.
386*
**1.
429*
**(0
.015
4)(0
.015
3)(0
.024
0)
Tim
e-Pr
oduc
tFE
No
Yes
Yes
Des
tina
tion
FEN
oN
oYe
s
Obs
erva
tion
s14
1920
1419
2014
1920
R-s
quar
ed0.
027
0.12
00.
234
Not
es:
Col
umns
1th
roug
h3
regr
ess
log
dist
ance
onpr
oduc
tav
aila
bilit
yat
the
mon
thly
Tim
ePe
riod
-Pro
duct
-Loc
atio
nle
vel
byor
dina
ryle
ast
squa
res.
Sam
ple
rest
rict
edto
prod
uct-
loca
tion
pair
sfo
rw
hich
the
prod
ucti
sob
serv
edin
atle
asto
nem
onth
.C
olum
n2
incl
udes
prod
uct-
mon
thfix
edef
fect
s.C
olum
n3
incl
udes
both
prod
uct-
mon
thfix
edan
dde
stin
atio
nfix
edef
fect
s.Ti
me-
prod
uct
clus
tere
dst
anda
rder
rors
inpa
rent
hese
s.*
sign
ifica
ntat
10pe
rcen
tle
vel,
**at
5pe
rcen
tan
d**
*at
1pe
rcen
t.
60
Tabl
e12
:Reg
ress
ing
Prod
uctA
vaila
bilit
yon
Dis
tanc
e:N
iger
ia
(1)
(2)
(3)
Prod
uctA
vaila
bilit
yat
Leve
lofT
ime-
Prod
uct-
Loca
tion
(1=P
rice
Rec
orde
d,2=
No
Pric
eR
ecor
ded)
Log
dist
ance
to-0
.040
7***
-0.0
468*
**-0
.013
1so
urce
(min
utes
)(0
.004
84)
(0.0
0446
)(0
.020
7)
Con
stan
t1.
011*
**1.
049*
**0.
939*
**(0
.031
1)(0
.027
7)(0
.127
)
Tim
e-Pr
oduc
tFE
No
Yes
Yes
Des
tina
tion
FEN
oN
oYe
s
Obs
erva
tion
s31
436
3143
631
436
R-s
quar
ed0.
004
0.28
20.
384
Not
es:
Col
umns
1th
roug
h3
regr
ess
log
dist
ance
onpr
oduc
tav
aila
bilit
yat
the
mon
thly
Tim
ePe
riod
-Pro
duct
-Loc
atio
nle
vel
byor
dina
ryle
ast
squa
res.
Sam
ple
rest
rict
edto
prod
uct-
loca
tion
pair
sfo
rw
hich
the
prod
ucti
sob
serv
edin
atle
asto
nem
onth
.Col
umn
2in
clud
espr
oduc
t-m
onth
fixed
effe
cts.
Col
umn
3in
clud
esbo
thpr
oduc
t-m
onth
fixed
and
dest
inat
ion
fixed
effe
cts.
Tim
e-pr
oduc
tcl
uste
red
stan
dard
erro
rsin
pare
nthe
ses.
*si
gnifi
cant
at10
perc
ent
leve
l,**
at5
perc
ent
and
***
at1
perc
ent.
61