https://orcid.org/0000-0002-6469-449X
Abstract
We use the microdata underlying the Ethiopian consumer price index to examine the spatial dispersion in local prices and availability of 401 items across 106 cities. Remote cities face higher prices and have access to fewer products. Large cities also face higher individual prices but enjoy access to a wider set of products. To assess the welfare implications of these patterns, we aggregate the data and build spatial cost-of-living indexes that account for both the price of available products and product availability. The cost of living is higher in remote and large cities. Moving from the first to the ninth decile in terms of remoteness (holding population size constant) results in an 8.3 per cent increase in the cost of living. A comparable move in terms of population size (holding remoteness constant) leads to a 3.7 per cent increase in the cost of living.
1. Introduction
How do prices and the availability of products vary across cities in developing countries? This question is important because spatial inequality is pervasive within poor countries, and accounting for cost-of-living differences across locations is key for a comprehensive view of these patterns.1 The literature in economic geography emphasizes city size as a key determinant not only of the dispersion in prices but also of product availability across cities (Handbury and Weinstein 2015; Feenstra, Xu, and Antoniades 2020). A key insight from these papers is that product availability increases with city size and greatly affects cost-of-living measures across regions. While they focus on the USA and China, product availability is also likely to matter in developing countries (see, e.g., FAO, IFAD, WHO WFP, and UNICEF 2018; WFP and CSA 2018; FAO, IFAD, WHO WFP, and UNICEF 2019). The literature in development economics often examines price differences between urban and rural areas (see, e.g., Ravallion and Van De Walle 1991; Deaton and Tarozzi 2005; Muller 2002), and highlights the role of transport costs in determining the price of available products in remote destinations (see, e.g., Rancourt, Bellavance, and Goentzel 2014; Atkin and Donaldson 2015). Our goal is to bridge these two strands of the literature and analyze how both city size and remoteness shape the spatial dispersion in the level of prices and the availability of a large number of products that are representative of households’ consumption in the context of a large developing country—Ethiopia.
We leverage the microdata underlying the Ethiopian consumer price index (CPI) to examine the spatial dispersion in local prices and product availability across more than 100 cities. Our findings indicate that individual prices tend to increase with both city size and remoteness. Furthermore, product availability, defined as the probability of finding a product in a given city, increases with city size but decreases with the level of remoteness. These results remain robust even after controlling for potential confounding factors such as income per capita, the prevalence of home production, and the proximity to international trade corridors. They also resist an IV strategy and various sample restrictions. The analysis at the micro-level shows that overall, remoteness has a clear impact on the local cost of living by driving up prices and reducing product availability, while the advantage of living in a larger city is ambiguous.
To quantify the impact of these micro-patterns for local consumers, we aggregate the data and calculate local cost-of-living indexes that capture variations in both price levels and product availability. The elasticity of substitution is the key parameter governing the utility cost of missing products in these spatial price indexes: the lower this elasticity, the more costly missing products are. We calibrate the elasticity using sectoral estimates by Broda, Greenfield, and Weinstein (2017) for ten African countries.
The analysis reveals that remote locations are unambiguously more expensive than central cities. Comparing cities at the first and ninth deciles in terms of remoteness, the price of available products is 5.2 per cent higher in remote cities. Furthermore, remote cities suffer less available products. When both margins are accounted for, the local cost-of-living is 8.3 per cent higher in remote cities.
The effect of city size on the local price index is more ambiguous and depends on the calibrated elasticity of substitution. Available products are 11 per cent more expensive in large cities relative to small ones (still based on the interdecile comparison). However, more products are available in large cities. Depending on the elasticity of substitution between these products, the greater availability of products in large cities may compensate for the higher individual prices. Based on our calibration, the cost-of-living that accounts for both margins remains 3.7 per cent higher in large cities. Note that the greater product availability in large cities exactly compensates for the higher price of available products for an elasticity of 3.
All in all, this quantification shows that when both the price of available products and product availability are accounted for, remoteness imposes a higher toll on the local cost-of-living than city-size.
1.1 Theoretical mechanisms
From a theoretical viewpoint, there are multiple mechanisms that may govern the relationship between product prices and product availability on the one hand, and city size and remoteness on the other. Regarding the relationship between city remoteness and prices, several factors are at play. First, more remote locations, almost by definition, suffer higher trade costs (including transport costs, time costs, and/or administrative costs), which should result into higher prices with a magnitude that depends on the pass-through of trade costs to final prices (Martin 2012). Second, according to the well-known Alchian–Allen effect, when trade costs are (at least partly) additive, only higher-quality varieties will make it into remote locations (see, e.g., Hummels and Skiba 2004; Martin and Mayneris 2015). Third, competition in remote locations may be lower and, in case of imperfect competition, this can lead to higher markups and thus higher prices (see, e.g., Asturias, García-Santana, and Ramos 2019). Hence, it makes almost no doubt that prices should be higher in remote locations. Moreover, due to the higher trade costs associated with remoteness (whether due to the distance, the quality of transport infrastructure, or in a developing country context the safety of the environment), we also unambiguously expect that more remote locations suffer more unavailable products/varieties.
The relationship between city size and prices is more ambiguous. On the one hand, competition tends to increase with city size which should induce, when markups are variable, lower prices in larger cities (Melitz and Ottaviano 2008; Feenstra, Xu, and Antoniades 2020). On the other hand, production and distribution costs may be higher in larger cities due to higher wages and higher real estate prices (see, e.g., Combes, Duranton, and Gobillon 2008, 2019), leading to higher prices for consumers. The net effect of both mechanisms is indeterminate.
Regarding the relationship between city size and product availability, models featuring consumers’ love for variety and size differences across regions predict that the number of available varieties increases with market size (Krugman 1991; Ottaviano, Tabuchi, and Thisse 2002), which may lead to self-reinforcing agglomeration mechanisms. We thus unambiguously expect a negative correlation between city size and the probability of a product or a variety being unavailable.
We provide exploratory analyses on some of these mechanisms but leave a complete assessment for future research.
1.2 Related literature
We contribute to three strands of the literature. First, we contribute to the literature on spatial differences in terms of the cost of living in developing countries (see, e.g., Ravallion and Van De Walle 1991; Deaton and Tarozzi 2005; Muller 2002; Timmins 2006; Ferré, Ferreira, and Lanjouw 2012). Our work offers three important contributions: (1) Our measure of the cost of living takes into account both the availability of products and services, as well as their prices when available, which distinguishes it from existing papers that typically focus solely on prices. Our results suggest that the cost of living is higher in large cities despite the presence of a broader set of available varieties. This may partly explain the higher wage premia in large urban cities discussed in the literature (Gollin, Kirchberger, and Lagakos 2021). (2) By leveraging the micro-data underlying the CPI in Ethiopia, our analysis provides broader spatial and industrial coverage compared to existing studies that often focus on a narrower range of products and specific villages or regions. (3) Instead of emphasizing the urban-rural divide, we analyze differences in the cost of living across cities and examine their relationship with economic geography factors, namely, population size and geographic remoteness of cities. Whereas we focus on differences in cost-of-living across space within a country, a related literature investigates how cost-of-living varies across income groups (see, e.g., Hottman and Monarch 2020; Faber and Fally 2022; Ma et al., 2024).
Second, a couple of recent papers examine prices and product availability in the context of Ethiopia (Krishnan and Zhang 2020; Gunning, Krishnan, and Mengistu 2024). Gunning, Krishnan, and Mengistu (2024) demonstrate that households in remote villages of Ethiopia have access to a lower variety of goods. Our work differs from these other papers along two important dimensions. First, the focus is different. These papers examine individual prices and product availability in remote villages. Our analysis instead covers Ethiopian main cities. Second, our data allow us to compute local price indexes and thus to directly compare the welfare across cities—and relate welfare differences to city size and remoteness.2
Third, our article contributes to the literature on the measurement and determinants of price indexes in the presence of missing products. Seminal papers in this literature have developed methods to measure the costs and benefits of disappearing and appearing varieties over time (Feenstra 1994; Hausman 1996; Diewert and Feenstra 2019). A few papers apply this method to a spatial context, and explore the link between the cost of living and city size (see, e.g., Handbury and Weinstein 2015; Feenstra, Xu, and Antoniades 2020, using US and Chinese data, respectively).3 We complement this literature by applying the method to a developing country for the first time and showing that, in the case of Ethiopia, remoteness has a stronger impact on spatial variations in terms of cost of living than city size.
Three papers are particularly relevant to our analysis. Atkin and Donaldson (2015) use the same source of data as ours to estimate the size of intra-national trade costs and their pass-through in the price paid by final consumers. To implement their methodology, they have to focus on fifteen products (out of 400) that are as precise as barcode categories and for which they can identify, based on interviews, the location of production. They find that internal trade costs are high in Ethiopia, and that falling trade barriers only modestly benefit final consumers as most of the gains are captured by intermediaries. Handbury and Weinstein (2015) use barcode data for food products in the USA and compute city-level measures of cost-of-living that account for both the availability and the price of varieties. They find that food products are more expensive in large cities because consumers there buy more expensive varieties in more expensive retail stores. On the other hand, more varieties are available in large cities. When both product availability and store heterogeneity are accounted for, large US cities exhibit a lower price index for food products. Focusing on nineteen grocery barcode products, Feenstra, Xu, and Antoniades (2020) show that in China, more varieties are available in large cities (as in the USA) and sold at a lower price (contrary to the USA). They attribute this latter pro-competitive effect to a more uneven distribution of manufacturers’ sales and retailers across space compared to the USA. The cost-of-living is then unambiguously lower in large cities in China.
Our work differs from these three papers in two main aspects. Firstly, we examine the impact of city size and remoteness on the availability and price of products across space. To our knowledge, we are the first to consider all these dimensions simultaneously. Second, we rely on the price data used for the construction of the Ethiopian CPI rather than barcode data. The main advantage of such data is that it allows us to study a basket of goods and services representative of the consumption of households. The limit is that we compare the price of similar but not necessarily identical products. Barcode data instead allow for the comparison of identical products but for a limited set of food (and dry grocery) products. Barcode data are not available in the context of developing countries like Ethiopia. Put differently, our approach compares, across cities, the typical prices faced by a consumer that wants to buy a given product, while approaches relying on barcode data compare the prices faced by a consumer willing to buy the exact same variety within a given retail chain. These two approaches correspond to different thought experiments.
The rest of the article is organized as follows. Section 2 presents the data, Section 3 discusses the results at the level of individual products, and Section 4 proposes an aggregation procedure to analyze spatial differences in terms of cost-of-living indexes. Finally, Section 5 concludes.
2. Data
2.1 Prices and product availability
The price data we use are those collected by the Ethiopian Central Statistical Agency (CSA) to construct the national CPI. A detailed description of the data extraction and treatment procedure is provided in the Supplementary Appendix. Prices are collected on a monthly basis by enumerators in 117 markets. In the original dataset, Addis Ababa is divided into twelve markets. To consolidate the data, we merge these markets into a single entity by considering the median price across the 12 markets, resulting in a total of 106 markets. Within each market, enumerators survey a predetermined sample of outlets every day during the first two weeks of each month. These outlets encompass a representative selection of open markets, kiosks, groceries, butcheries, pharmacies, supermarkets, and other similar establishments. Enumerators are instructed to locate specific products and report them as missing if they are unable to find them. When a product is found, enumerators ascertain its typical price (after bargaining) by conducting interviews with both sellers and consumers (Atkin and Donaldson 2015).
The survey encompasses 427 products and services, which can be categorized into 12 major groups and 55 categories. These items include food products like bread and cereals, as well as clothing and footwear, household equipment, and services such as haircuts and restaurants. The product descriptions range from specific branded items (e.g., “Coca-Cola bottle 300c”) to more detailed products without a specific brand (e.g., “bed sheet (Patterned Kombolcha) 1.90 m × 2.50 m”) and even generic product categories (e.g., “sorghum yellow, kg” or “rice imported, kg”). The index covers all type of expenditures including those related to housing such as stone and sand for construction or concrete blocks. However, there is no information on rents in our dataset. Another dataset provided by the Ethiopian Statistical Agency reports that rents account for only 3.3 per cent of expenditures across Ethiopian regions (the maximum being 10 per cent in Addis Ababa). We are thus confident we cover the lion share of consumers’ expenditures with the data we use.
To focus on location-based price differences rather than changes over time, and to address potential issues related to misreporting, we use the monthly data for the year 2015 and calculate the median price per product and location based on the months when the product is available.4 In addition to price information, the survey allows us to identify unavailable products. We consider a product to be unavailable in a given month if the price is missing (Atkin and Donaldson 2015, do the same for example). Then we require the product to be missing every month throughout 2015 to be classified as unavailable in that year. After excluding products reported in less than 10 cities, the final dataset contains information on 401 products across 106 cities, that is, 42,506 product-market pairs, out of which approximately 34 per cent are categorized as missing. Missing is more prevalent for some services related to education (45 per cent), communication (49 per cent), and for transport including airplane, bus etc. (62 per cent). Two categories stand out with very few missing observations: health including drugs (15 per cent) and restaurants, bars, and hotels (8 per cent).
Because inflation is a significant political issue in Ethiopia, there may be concerns about potential manipulation of price-quote data for political reasons. However, it is worth noting that 2015 was not a year of hyperinflation, which limits the political motivations for price manipulation. To further evaluate the reliability of the price quotes, we test whether the data adhere to Benford’s Law. Researchers have used deviations from this law to identify reporting issues in survey data (e.g., Judge and Schechter 2009; Demir and Javorcik 2020). The underlying concept of this test is that manipulating the data while still conforming to the Benford distribution of the first digit of numerical data is difficult. Supplementary Appendix Fig. A.2 presents the frequency distribution of the first digit of price quotes in our data alongside the expected distribution according to Benford’s Law. The observed frequencies align well with the frequencies predicted by the Benford law.5 Hence, the distribution of the first digit of price quotes in the data is consistent with Benford’s Law, indicating that price manipulation is not a major concern in this context.
2.2 Size and remoteness
Our objective is to assess the impact of city size and remoteness on the spatial dispersion of the cost of living. City size is measured by population, and the population data used are based on the 2007 population and housing census of Ethiopia, with population projection figures for 2015 provided by the Ethiopian CSA (see CSA 2013). However, for 25 per cent of the urban centers, the CSA does not offer any projections. In such cases, we rely on the projection figures for 2015 provided by the Ethiopian Ministry of Water and Energy in 2011, as part of the urban water-supply universal access plan (see Ministry of Water and Energy 2011). The city sizes in the dataset range from 764 inhabitants (Deri) to 3,273,000 (Addis Ababa). Among the cities in the sample, 63 per cent have a population of fewer than 30,000 inhabitants, and these cities are classified as “rural.”
One might wonder how large or remote cities differ from other locations. First, the correlation between city size and remoteness is negative but not significantly different from zero in the data. Similarly, ethnic diversity is not correlated with either city size or remoteness. Both large and remote cities tend to have a dominant ethnicity that is significantly less connected to the ethnicity of the prime minister. Last, remote regions are far from Addis, while cities nearby international corridors tend to be bigger in terms of population (see Supplementary Appendix Table A.5).
2.3 Other data sources
The baseline specification links observed prices and product availability to city size and remoteness. We check the robustness of estimated relationships to possible confounding factors including income per capita, ethnic diversity, and home production. The construction of price indices further necessitates information on expenditure weights. The construction of each of these variables is described in the Supplementary Appendix.
3. Spatial dispersion of products’ prices and availability
This section examines the link between the price and availability of individual items and the size and remoteness of cities.
3.1 Empirical approach
The purpose of this regression analysis on individual data is mainly descriptive, as the exact value of the coefficients we obtain on the various covariates has no importance for the analysis of city-level price indices we conduct in the last part of the article. We nonetheless check the robustness of our results to two econometric concerns: endogeneity of remoteness and selection bias. Endogeneity may arise in our specification if factors influencing travel time between cities also influence the level of prices and the availability of products. For example, local economic conditions, local conflicts, or connections between local politicians and the central government may all affect the quantity and the quality of road infrastructure on the one hand, and products’ prices and availability on the other. To tackle this issue, we first include additional controls such as, among others, local ethnic diversity, distance to trade corridors, and home production. We then propose an IV strategy where remoteness is instrumented by the average bilateral geodesic distance to other cities.
Selection bias may arise in our price regression. Indeed, we do not observe the price when a product is missing, which may imply that the sample of cities for which prices are observed is not representative. For instance, if some observations are missing because a product becomes prohibitively expensive due to remoteness or city size, we may underestimate the impact of these variables if we don’t account for selection bias. We develop two strategies to tackle this issue. First, we check the robustness of our results on a sample of products that are available across almost all location. Second, we implement a Heckman model of sample selection.
3.2 Baseline results
Columns (1)–(4) of Table 1 present the estimates of the relationship between prices and remoteness and city size. As can be seen from column (1), remoteness and city size are significantly related to the local prices of available products. Large and more remote cities are more expensive than the others. Individual prices are 11.6 per cent higher in a city of 159,300 inhabitants (p90 in the sample) than in a city with 8,685 inhabitants (p10 in the sample). Individual prices are 5.9 per cent higher in remote cities than in more central ones (again p90 versus p10 of the remoteness index in the sample). These results are robust to controlling for per capita income (column (2)), measuring remoteness using a population-weighted average of distance to other cities (column (3), P-value on population-weighted remoteness equal to 11 per cent), and excluding Addis Ababa from the sample (column (4)).
Price and availability of products across Ethiopia.
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | |
|---|---|---|---|---|---|---|---|---|
| Price of individual product (log) | Product not available (dummy) | |||||||
| Ln Remoteness | 0.119 | 0.112 | 0.122 | 0.105 | 0.100 | 0.116 | ||
| (0.032) | (0.036) | (0.034) | (0.061) | (0.065) | (0.061) | |||
| Ln Population | 0.040 | 0.044 | 0.039 | 0.038 | −0.072 | −0.068 | −0.072 | −0.075 |
| (0.004) | (0.005) | (0.004) | (0.005) | (0.007) | (0.009) | (0.007) | (0.008) | |
| Ln per cap. inc. | −0.025 | −0.022 | ||||||
| (0.022) | (0.032) | |||||||
| Ln pop-weighted | 0.041 | 0.058 | ||||||
| remoteness | (0.026) | (0.044) | ||||||
| Addis | Yes | Yes | No | Yes | Yes | Yes | Yes | No |
| Observations | 28,183 | 28,183 | 28,183 | 27,810 | 42,506 | 42,506 | 42,506 | 42,105 |
| R-squared | 0.03 | 0.03 | 0.02 | 0.03 | 0.06 | 0.06 | 0.06 | 0.06 |
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | |
|---|---|---|---|---|---|---|---|---|
| Price of individual product (log) | Product not available (dummy) | |||||||
| Ln Remoteness | 0.119 | 0.112 | 0.122 | 0.105 | 0.100 | 0.116 | ||
| (0.032) | (0.036) | (0.034) | (0.061) | (0.065) | (0.061) | |||
| Ln Population | 0.040 | 0.044 | 0.039 | 0.038 | −0.072 | −0.068 | −0.072 | −0.075 |
| (0.004) | (0.005) | (0.004) | (0.005) | (0.007) | (0.009) | (0.007) | (0.008) | |
| Ln per cap. inc. | −0.025 | −0.022 | ||||||
| (0.022) | (0.032) | |||||||
| Ln pop-weighted | 0.041 | 0.058 | ||||||
| remoteness | (0.026) | (0.044) | ||||||
| Addis | Yes | Yes | No | Yes | Yes | Yes | Yes | No |
| Observations | 28,183 | 28,183 | 28,183 | 27,810 | 42,506 | 42,506 | 42,506 | 42,105 |
| R-squared | 0.03 | 0.03 | 0.02 | 0.03 | 0.06 | 0.06 | 0.06 | 0.06 |
Notes: The dependent variables are the log price (columns (1)–(5)) and a dummy equal to 1 if the product is unavailable (columns (6)–(10)). These dependent variables are defined in the dimension. Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and , respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels.
Price and availability of products across Ethiopia.
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | |
|---|---|---|---|---|---|---|---|---|
| Price of individual product (log) | Product not available (dummy) | |||||||
| Ln Remoteness | 0.119 | 0.112 | 0.122 | 0.105 | 0.100 | 0.116 | ||
| (0.032) | (0.036) | (0.034) | (0.061) | (0.065) | (0.061) | |||
| Ln Population | 0.040 | 0.044 | 0.039 | 0.038 | −0.072 | −0.068 | −0.072 | −0.075 |
| (0.004) | (0.005) | (0.004) | (0.005) | (0.007) | (0.009) | (0.007) | (0.008) | |
| Ln per cap. inc. | −0.025 | −0.022 | ||||||
| (0.022) | (0.032) | |||||||
| Ln pop-weighted | 0.041 | 0.058 | ||||||
| remoteness | (0.026) | (0.044) | ||||||
| Addis | Yes | Yes | No | Yes | Yes | Yes | Yes | No |
| Observations | 28,183 | 28,183 | 28,183 | 27,810 | 42,506 | 42,506 | 42,506 | 42,105 |
| R-squared | 0.03 | 0.03 | 0.02 | 0.03 | 0.06 | 0.06 | 0.06 | 0.06 |
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | |
|---|---|---|---|---|---|---|---|---|
| Price of individual product (log) | Product not available (dummy) | |||||||
| Ln Remoteness | 0.119 | 0.112 | 0.122 | 0.105 | 0.100 | 0.116 | ||
| (0.032) | (0.036) | (0.034) | (0.061) | (0.065) | (0.061) | |||
| Ln Population | 0.040 | 0.044 | 0.039 | 0.038 | −0.072 | −0.068 | −0.072 | −0.075 |
| (0.004) | (0.005) | (0.004) | (0.005) | (0.007) | (0.009) | (0.007) | (0.008) | |
| Ln per cap. inc. | −0.025 | −0.022 | ||||||
| (0.022) | (0.032) | |||||||
| Ln pop-weighted | 0.041 | 0.058 | ||||||
| remoteness | (0.026) | (0.044) | ||||||
| Addis | Yes | Yes | No | Yes | Yes | Yes | Yes | No |
| Observations | 28,183 | 28,183 | 28,183 | 27,810 | 42,506 | 42,506 | 42,506 | 42,105 |
| R-squared | 0.03 | 0.03 | 0.02 | 0.03 | 0.06 | 0.06 | 0.06 | 0.06 |
Notes: The dependent variables are the log price (columns (1)–(5)) and a dummy equal to 1 if the product is unavailable (columns (6)–(10)). These dependent variables are defined in the dimension. Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and , respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels.
Columns (5)–(8) display the results on product availability. From column (5), we see that the probability that a product is missing is significantly higher in remote cities and significantly lower in large cities. Still using the first and ninth deciles of the distribution in terms of population size and remoteness, the probability that a product is unavailable is 21 percentage points higher in small cities than in big cities, and 5.2 percentage points higher in remote cities than in central ones. The results are overall robust to controlling for income per capita (column (6), P-value on remoteness equal to 12 per cent in this case) and excluding Addis Ababa (column (8)). The impact of remoteness remains positive but not statistically significant with an alternative population-weighted measure of distance to other cities (column (7)). This suggests that the distance to other markets, independent of their size, is the primary impediment to product availability in a city.
3.3 Sensitivity analysis
3.3.1 Sectoral and temporal heterogeneity
We run the benchmark regressions separately for the thirty-seven product categories used later in the article to build local price indexes. Our baseline results do not change much across product categories from a qualitative viewpoint, even though the value of the coefficients does vary. The coefficients are reported in Supplementary Appendix Table A.6.8
Note that we have also checked that our results are stable over years and across seasons (results are available upon request). We find that the effects of city size and remoteness on prices and product availability are stable if we consider prices for 2011–2015 (the years for which we manage to collect the data).9 We also used our monthly data in 2015 to examine how our estimates vary across seasons. We considered three seasons: a dry season (Bega) from October to January; a short rainy season (Belg) from February to May; and the main rainy season (Kiremt) from June to September. We do not find a significant heterogeneity across seasons except in the case of agricultural products for which the effect of remoteness on prices shrinks during the harvest season (Belg). One possible interpretation is that the supply of locally produced goods in remote areas reduces reliance on external supplies, leading to a temporary convergence in prices between remote and more accessible regions.
3.3.2 Additional controls
Table 2 examines the sensitivity of the results to the inclusion of several additional controls. Remoteness might be more related to the distance to the capital city, Addis Ababa, which happens to be close to the geographic center of the country too. Moreover, Ethiopia is a landlocked country and Atkin and Donaldson (2015) argue that 90 per cent of trade flows enter Ethiopia via Djibouti through a commercial corridor, whose main Ethiopian city is Kombolcha. We thus control for distance to these two cities in columns (1) and (3). None of these variables is significantly related to the price and availability of products in our benchmark sample.
Sensitivity analysis.
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) | |
|---|---|---|---|---|---|---|---|---|---|---|
| Price (log) | Missing | Price (log) | Missing | Price (log) | Missing | |||||
| Ln Remoteness | 0.079 | 0.114 | 0.087 | 0.082 | 0.175 | 0.176 | 0.167 | 0.128 | ||
| (0.033) | (0.062) | (0.040) | (0.039) | (0.042) | (0.042) | (0.033) | (0.073) | |||
| Ln Population | 0.039 | 0.035 | −0.068 | −0.071 | 0.046 | 0.038 | −0.047 | −0.044 | 0.040 | −0.072 |
| (0.005) | (0.004) | (0.007) | (0.007) | (0.006) | (0.007) | (0.005) | (0.006) | (0.004) | (0.007) | |
| Ln Travel time to Addis | 0.009 | 0.005 | ||||||||
| (0.006) | (0.010) | |||||||||
| Ln Travel time to import corridor | −0.010 | 0.017 | ||||||||
| (0.008) | (0.011) | |||||||||
| Dominant ethnicity politically connected | −0.051 | 0.012 | ||||||||
| (0.013) | (0.021) | |||||||||
| Log Ethnic diversity index | 0.006 | 0.004 | ||||||||
| (0.015) | (0.030) | |||||||||
| Home production | −0.089 | 0.023 | ||||||||
| (0.032) | (0.023) | |||||||||
| Observations | 27,697 | 28,183 | 41,704 | 42,506 | 8,672 | 8,672 | 12,370 | 12,370 | 28,183 | 42,506 |
| R-squared | 0.03 | 0.03 | 0.03 | 0.06 | 0.04 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 |
| Kleinbergen–Paap F-stat. | 627.33 | 663.58 | ||||||||
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) | |
|---|---|---|---|---|---|---|---|---|---|---|
| Price (log) | Missing | Price (log) | Missing | Price (log) | Missing | |||||
| Ln Remoteness | 0.079 | 0.114 | 0.087 | 0.082 | 0.175 | 0.176 | 0.167 | 0.128 | ||
| (0.033) | (0.062) | (0.040) | (0.039) | (0.042) | (0.042) | (0.033) | (0.073) | |||
| Ln Population | 0.039 | 0.035 | −0.068 | −0.071 | 0.046 | 0.038 | −0.047 | −0.044 | 0.040 | −0.072 |
| (0.005) | (0.004) | (0.007) | (0.007) | (0.006) | (0.007) | (0.005) | (0.006) | (0.004) | (0.007) | |
| Ln Travel time to Addis | 0.009 | 0.005 | ||||||||
| (0.006) | (0.010) | |||||||||
| Ln Travel time to import corridor | −0.010 | 0.017 | ||||||||
| (0.008) | (0.011) | |||||||||
| Dominant ethnicity politically connected | −0.051 | 0.012 | ||||||||
| (0.013) | (0.021) | |||||||||
| Log Ethnic diversity index | 0.006 | 0.004 | ||||||||
| (0.015) | (0.030) | |||||||||
| Home production | −0.089 | 0.023 | ||||||||
| (0.032) | (0.023) | |||||||||
| Observations | 27,697 | 28,183 | 41,704 | 42,506 | 8,672 | 8,672 | 12,370 | 12,370 | 28,183 | 42,506 |
| R-squared | 0.03 | 0.03 | 0.03 | 0.06 | 0.04 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 |
| Kleinbergen–Paap F-stat. | 627.33 | 663.58 | ||||||||
Notes: Dependent variables are defined in the dimension. The variable “Dominant ethnicity politically connected” is a dummy equal to one when the dominant ethnic group in the city is the same as the Prime Minister’s one. The index of ethnic diversity is the inverse of a Herfindahl index based on the share of the various ethnic groups in the population. Home production measures the share of auto-consumption in total households’ consumption by region and type of area (rural or urban). Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels. Columns (9) and (10) are the outcomes of IV regressions where Ln Remoteness (which measures the average travel time by road to all the other cities) is instrumented by Ln Average geodesic distance to all the other cities.
Sensitivity analysis.
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) | |
|---|---|---|---|---|---|---|---|---|---|---|
| Price (log) | Missing | Price (log) | Missing | Price (log) | Missing | |||||
| Ln Remoteness | 0.079 | 0.114 | 0.087 | 0.082 | 0.175 | 0.176 | 0.167 | 0.128 | ||
| (0.033) | (0.062) | (0.040) | (0.039) | (0.042) | (0.042) | (0.033) | (0.073) | |||
| Ln Population | 0.039 | 0.035 | −0.068 | −0.071 | 0.046 | 0.038 | −0.047 | −0.044 | 0.040 | −0.072 |
| (0.005) | (0.004) | (0.007) | (0.007) | (0.006) | (0.007) | (0.005) | (0.006) | (0.004) | (0.007) | |
| Ln Travel time to Addis | 0.009 | 0.005 | ||||||||
| (0.006) | (0.010) | |||||||||
| Ln Travel time to import corridor | −0.010 | 0.017 | ||||||||
| (0.008) | (0.011) | |||||||||
| Dominant ethnicity politically connected | −0.051 | 0.012 | ||||||||
| (0.013) | (0.021) | |||||||||
| Log Ethnic diversity index | 0.006 | 0.004 | ||||||||
| (0.015) | (0.030) | |||||||||
| Home production | −0.089 | 0.023 | ||||||||
| (0.032) | (0.023) | |||||||||
| Observations | 27,697 | 28,183 | 41,704 | 42,506 | 8,672 | 8,672 | 12,370 | 12,370 | 28,183 | 42,506 |
| R-squared | 0.03 | 0.03 | 0.03 | 0.06 | 0.04 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 |
| Kleinbergen–Paap F-stat. | 627.33 | 663.58 | ||||||||
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) | |
|---|---|---|---|---|---|---|---|---|---|---|
| Price (log) | Missing | Price (log) | Missing | Price (log) | Missing | |||||
| Ln Remoteness | 0.079 | 0.114 | 0.087 | 0.082 | 0.175 | 0.176 | 0.167 | 0.128 | ||
| (0.033) | (0.062) | (0.040) | (0.039) | (0.042) | (0.042) | (0.033) | (0.073) | |||
| Ln Population | 0.039 | 0.035 | −0.068 | −0.071 | 0.046 | 0.038 | −0.047 | −0.044 | 0.040 | −0.072 |
| (0.005) | (0.004) | (0.007) | (0.007) | (0.006) | (0.007) | (0.005) | (0.006) | (0.004) | (0.007) | |
| Ln Travel time to Addis | 0.009 | 0.005 | ||||||||
| (0.006) | (0.010) | |||||||||
| Ln Travel time to import corridor | −0.010 | 0.017 | ||||||||
| (0.008) | (0.011) | |||||||||
| Dominant ethnicity politically connected | −0.051 | 0.012 | ||||||||
| (0.013) | (0.021) | |||||||||
| Log Ethnic diversity index | 0.006 | 0.004 | ||||||||
| (0.015) | (0.030) | |||||||||
| Home production | −0.089 | 0.023 | ||||||||
| (0.032) | (0.023) | |||||||||
| Observations | 27,697 | 28,183 | 41,704 | 42,506 | 8,672 | 8,672 | 12,370 | 12,370 | 28,183 | 42,506 |
| R-squared | 0.03 | 0.03 | 0.03 | 0.06 | 0.04 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 |
| Kleinbergen–Paap F-stat. | 627.33 | 663.58 | ||||||||
Notes: Dependent variables are defined in the dimension. The variable “Dominant ethnicity politically connected” is a dummy equal to one when the dominant ethnic group in the city is the same as the Prime Minister’s one. The index of ethnic diversity is the inverse of a Herfindahl index based on the share of the various ethnic groups in the population. Home production measures the share of auto-consumption in total households’ consumption by region and type of area (rural or urban). Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels. Columns (9) and (10) are the outcomes of IV regressions where Ln Remoteness (which measures the average travel time by road to all the other cities) is instrumented by Ln Average geodesic distance to all the other cities.
Also, cities where the dominant ethnic group is the same as the Prime Minister’s one could benefit from lower prices and greater product availability thanks to political connections and favoritism. Another issue with ethnicity is that if ethnic groups have very specific tastes, more ethnically diverse cities could be cities where products are more likely to be available (e.g., Schiff 2015), and this diversity could also affect the price at which they are sold. We thus control for a dummy equal to one when the dominant ethnic group in the city is the same as the Prime Minister’s one and for the inverse of a Herfindahl index based on the share of the various ethnic groups in the population in columns (2) and (4). Cities where the dominant ethnic group is the same as the Prime Minister’s one enjoy lower prices but not greater product availability. Ethnic diversity has no significant relationship with the price and availability of products. Both remoteness and city size remain significantly related to prices and product availability.
3.3.3 Home production
In the context of a developing country such as Ethiopia, another important issue that could undermine the benchmark results is home production. The phenomenon is particularly important in rural areas. If products are produced directly by those who consume them, they might be unavailable on the market without necessarily being unavailable for consumption. For a subset of food products in the database, we have information on the share of auto-consumption in total households’ consumption by region and type of area (rural or urban). After reproducing in columns (5) and (7) the benchmark results for the subset of observations for which home-production is available, we directly introduce home-production in the regression. Column (6) shows that home-production pushes prices downward, which is consistent with the view that it increases competition. However, we do not find a significant relationship between home-production and product availability. Importantly, the sign, magnitude, and significance of the coefficients on remoteness and population size are not affected by the introduction of the home-production variable.
3.3.4 Endogeneity
There may be other confounders (e.g., conflicts, political connections) that affect both remoteness and the dependent variables, and that are not accounted for by the previous exercises. To push further the causal analysis of remoteness, we instrument it by the average geodesic distance to all the other cities, capturing the effect of remoteness related to the relative physical location of cities, and not to other geo-political or socio-economic factors. The results displayed in columns (9) and (10) of Table 2 show that if anything, the coefficient on remoteness is boosted (and the one on population remains unaffected).
3.3.5 Sample selection bias
Focusing on the price of available products, we treat potential sample selection issues in Table 3. To deal with sample selection, we first replicate the benchmark regression on the subsamples of products for which the share of cities where the product is missing is alternatively below 5 per cent, 3 per cent, and 1 per cent. We find that the coefficient on population is stable across specifications. The coefficient on remoteness increases when we focus on samples of products which are less likely to be missing, which is coherent with the idea that products tend to be missing in remote locations when they would become prohibitively expensive if they were available.
Price and availability of products across Ethiopia.
| Benchmark | 5% missing | 3% missing | 1% missing | Heckman | |
|---|---|---|---|---|---|
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.119 | 0.117 | 0.136 | 0.191 | 0.127 |
| (0.035) | (0.032) | (0.059) | (0.038) | (0.041) | |
| Ln Population | 0.040 | 0.024 | 0.022 | 0.030 | 0.080 |
| (0.004) | (0.004) | (0.004) | (0.005) | (0.009) | |
| Observations | 28,183 | 8,529 | 6,802 | 3,801 | 28,183 |
| Benchmark | 5% missing | 3% missing | 1% missing | Heckman | |
|---|---|---|---|---|---|
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.119 | 0.117 | 0.136 | 0.191 | 0.127 |
| (0.035) | (0.032) | (0.059) | (0.038) | (0.041) | |
| Ln Population | 0.040 | 0.024 | 0.022 | 0.030 | 0.080 |
| (0.004) | (0.004) | (0.004) | (0.005) | (0.009) | |
| Observations | 28,183 | 8,529 | 6,802 | 3,801 | 28,183 |
Notes: The dependent variables are the log price defined in the dimension. Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels. In columns (2)–(4) we reduce the sample to products that are missing in less than 5 per cent, 3 per cent, and 1 per cent of cities.
Price and availability of products across Ethiopia.
| Benchmark | 5% missing | 3% missing | 1% missing | Heckman | |
|---|---|---|---|---|---|
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.119 | 0.117 | 0.136 | 0.191 | 0.127 |
| (0.035) | (0.032) | (0.059) | (0.038) | (0.041) | |
| Ln Population | 0.040 | 0.024 | 0.022 | 0.030 | 0.080 |
| (0.004) | (0.004) | (0.004) | (0.005) | (0.009) | |
| Observations | 28,183 | 8,529 | 6,802 | 3,801 | 28,183 |
| Benchmark | 5% missing | 3% missing | 1% missing | Heckman | |
|---|---|---|---|---|---|
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.119 | 0.117 | 0.136 | 0.191 | 0.127 |
| (0.035) | (0.032) | (0.059) | (0.038) | (0.041) | |
| Ln Population | 0.040 | 0.024 | 0.022 | 0.030 | 0.080 |
| (0.004) | (0.004) | (0.004) | (0.005) | (0.009) | |
| Observations | 28,183 | 8,529 | 6,802 | 3,801 | 28,183 |
Notes: The dependent variables are the log price defined in the dimension. Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels. In columns (2)–(4) we reduce the sample to products that are missing in less than 5 per cent, 3 per cent, and 1 per cent of cities.
We also implement a Heckman sample selection procedure. In the absence of any credible excluded variable that would affect the availability of products, but not their prices, the procedure relies on the functional form only (the first stage being non-linear). Again the results are qualitatively similar but the coefficients on remoteness and population are higher with the sample correction (especially for city size).
3.4 Mechanisms behind the price effects
Price differences across locations may reflect differences in transport costs, in quality, in distribution costs, or in markups. Whereas a quantitative assessment of these mechanisms is beyond the scope of the article, we have performed a series of exploratory analyses to shed light on them. The precise description of these analyses appears in the Supplementary Appendix. Here is a summary of our findings.
3.4.1 Prices and remoteness
First, we investigate whether the price premium paid in remote location reflects that remote locations are farther away from places where goods are produced. To do so, we focus on imported products for which we can assess the travel time between the main Ethiopian point of entry, Kombolcha (the transit point from Djibouti), and cities. The results show that cities that are farther away from the trade corridor exhibit higher prices for imported products, and that while the coefficient on remoteness slightly decreases, it remains positively and highly significantly related to the price of available varieties. This suggests that remoteness captures factors beyond travel time for shipping goods. One possible explanation is that there is less competition in remote cities, which pushes prices up. For a given travel time, transportation costs might also be higher if risk is higher on these roads due to conflicts for instance, or if there is less competition among carriers.
3.4.2 Prices and city size
Then, we explore different channels that may explain the price premium in large cities. A first explanation is that price differences reflects difference in terms of the quality of goods sold. We performed our analysis on a subsample of barcode products available in our dataset. In this (tiny) sample, the price premium vanishes, which suggests that part of the price premium in large cities may be explained by composition effects.
To make further progress on this issue, we also examine whether the elasticity of prices to population size varies with the length of the quality ladder of products. Specifically, we define products with large cross-city price variations as more likely to exhibit quality differences. We find that the elasticity of prices to city size is higher for products potentially affected by vertical differentiation. Together, the results suggest that quality variations across cities may partially—but not entirely—explain the observed correlation between city size and prices.
An alternative explanation is that distribution costs are higher in large cities, which increases the price paid by consumers.10 We leverage our dataset to compute proxies for local unskilled labor costs and local real-estate costs that may both affect distribution costs. Once we add these two proxies in our baseline specification, the price premium related to city size drops by 40 per cent. One interpretation of this result is that higher distribution costs are passed through to consumers, partially explaining why prices are generally higher in larger cities.
There is no smoking gun, but the analysis suggests that part of the price premium in remote locations is due to higher transport costs, and part of the price premium in large cities is driven by higher quality and higher distribution costs. However, individual prices collected by enumerators reflect, for a given product, the typical price of available varieties within a market. We believe these cross-city differences in prices are informative of spatial difference in terms of cost-of-living, irrespective of the mechanisms underlying them. This is why we keep using observed prices, and not prices purged of quality differences and/or distribution costs, in the rest of the analysis.
4. CPI across Ethiopian cities
So far, the results unambiguously suggest that remote cities are more expensive than central ones due to higher prices and lower product availability. On the other hand, prices are higher in large cities but consumers have access to a wider range of products. To our knowledge, we are the first to document this tension between product availability and individual prices in the context of a developing country. In this section, we put more structure and compute a spatial price index to assess which force dominates.
4.1 Spatial CES price index
To summarize, the price index is a weighted average of g-category sub-indices that have two components: an intensive component that is a standard price index that tracks the price gap across products in location c compared to a location of reference; and an extensive part that measures the utility cost of unavailable products in location c compared to the same reference location. Here, we assume the reference location is a fictitious city where all the products of all the product categories are available at a price equal to the median price observed in 2015 across Ethiopian cities in the sample.
We have information on expenditures across regions for fifty-five categories of products. Within these product categories, we assume that consumption is equally split across products. The formula of SPI is valid if the set of available products from product category g in city c is not empty. However, many cities in the sample have empty sets for some categories of products. We thus group these categories into 37 g groups. Details on these groups are provided in the Supplementary Appendix. This aggregation allows us to compute the formula for 78 out of 106 cities in the sample. 12
A key term in the formula is the elasticity of substitution , which affects the exact price index through its extensive part . If products are poor substitutes, then missing products are more costly for consumers. With the data at hand, we cannot directly compute . We thus use the elasticities estimated by Broda, Greenfield, and Weinstein (2017) based on international trade data at the three-digit level of the Harmonized System (HS) nomenclature for ten African countries (Algeria, Central African Republic, Egypt, Gabon, Madagascar, Malawi, Mauritius, Morocco, Togo, and Tunisia). We manually build a correspondence between the thirty-seven product groups of the Ethiopian price data and the HS three-digit nomenclature, and we propose two calibrations for . For the first one, we use the median of the sectoral estimates for the ten African countries available in the dataset of Broda, Greenfield, and Weinstein (2017). Since the estimates are based on trade in goods data, we have no information on the value of the substitution elasticity in services sectors (e.g., transportation, restaurants and cafes, personal care). For these sectors, we calibrate the elasticity to 3.37 (the median elasticity across goods estimated for African countries in Broda, Greenfield, and Weinstein 2017). In the end, across the Ethiopian product categories used to build the local price indexes, the median and the mean of the calibrated substitution elasticity are respectively equal to 3.37 and 5.08. In an alternative calibration, we do the same, but we restrict the sample of countries to Madagascar and Togo, which are the closest to Ethiopia in terms of income per capita and linguistic proximity. In this case, the substitution elasticity for services is set to 3.99 (i.e., the median across HS 3-digit sectors observed for the two countries).
4.2 Results
We first propose a visual inspection of the relationship between the cost of living as measured by the spatial price index and its various components on the one hand, and city size and remoteness on the other. The graphs in Fig. 1 plot the level of the various components of the price index against the degree of remoteness of the city, taking the calibration of substitution elasticities based on the 10 African countries in Broda, Greenfield, and Weinstein (2017); the size of the circle is proportional to the population of the city it represents. Panel (a) relates to the intensive component of the CES-price index , that is how expensive available products are compared to the median price quote observed in Ethiopia. Two main messages emerge: once expenditure shares are accounted for, large cities are still more expensive (big circles are at the top of the graph), and remote cities also still tend to exhibit higher prices (the slope of the scatter plot is positive). A notable exception is Addis Ababa whose price level is as high as the more remote cities. Regarding the extensive component of the price index in panel (b), when expenditure shares and the elasticity of substitution are taken into account, large and less remote cities appear with a lower value of the extensive component of the price index, which is coherent with the fact that the probability that a product is available is higher in large and less remote cities. The combination of the two components gives us the exact price index , which appears on panel (c). The relative effects of city size on the level of prices and on product availability vary with the value of . With the current calibration, the graph shows that the correlation of the exact price index with city size is positive, which means that the access to a wider range of products in large cities does not fully offset the price premium. Remote cities have a higher cost of living than central cities, which is intuitive since they suffer from both higher prices and lower product variety.
The geography of prices and product availability.
Notes: Each dot is an Ethiopian city. The size of the circle is proportional to the city’s population size. The log of the intensive part of the price index is in panel (a). The log of the extensive component is in panel (b). The exact price index in panel (c) is the sum of the indexes presented in panels (a) and (b). Indices presented in panels (b) and (c) are computed taking the calibration of substitution elasticities based on the ten African countries in Broda, Greenfield, and Weinstein 2017).
The graphical observations are largely confirmed by the econometric analysis reported in Table 4. The EPI is the product of the intensive and the extensive components, and thus log-linear in these two components.
City-level regressions-local CPI.
| Ln SPI | Ln VA | Ln EPI | Ln VA | Ln EPI | |
|---|---|---|---|---|---|
| calibrated using | calibrated using | ||||
| the 10 African countries in Broda, Greenfield, and Weinstein (2017) | Togo and Madagascar | ||||
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.103 | 0.062 | 0.165 | 0.041 | 0.144 |
| (0.037) | (0.029) | (0.050) | (0.024) | (0.048) | |
| Ln Population | 0.042 | −0.028 | 0.014 | −0.020 | 0.022 |
| (0.006) | (0.004) | (0.009) | (0.003) | (0.008) | |
| Observations | 78 | 78 | 78 | 78 | 78 |
| R-squared | 0.39 | 0.48 | 0.16 | 0.42 | 0.18 |
| Ln SPI | Ln VA | Ln EPI | Ln VA | Ln EPI | |
|---|---|---|---|---|---|
| calibrated using | calibrated using | ||||
| the 10 African countries in Broda, Greenfield, and Weinstein (2017) | Togo and Madagascar | ||||
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.103 | 0.062 | 0.165 | 0.041 | 0.144 |
| (0.037) | (0.029) | (0.050) | (0.024) | (0.048) | |
| Ln Population | 0.042 | −0.028 | 0.014 | −0.020 | 0.022 |
| (0.006) | (0.004) | (0.009) | (0.003) | (0.008) | |
| Observations | 78 | 78 | 78 | 78 | 78 |
| R-squared | 0.39 | 0.48 | 0.16 | 0.42 | 0.18 |
Notes: The dependent variable is the log spatial price index (Ln SPI) in column (1), the log spatial availability index (Ln VA) in columns (2) and (4), and the log exact price index (Ln EPI=Ln SPI + Ln VA) in columns (3) and (5). Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels.
City-level regressions-local CPI.
| Ln SPI | Ln VA | Ln EPI | Ln VA | Ln EPI | |
|---|---|---|---|---|---|
| calibrated using | calibrated using | ||||
| the 10 African countries in Broda, Greenfield, and Weinstein (2017) | Togo and Madagascar | ||||
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.103 | 0.062 | 0.165 | 0.041 | 0.144 |
| (0.037) | (0.029) | (0.050) | (0.024) | (0.048) | |
| Ln Population | 0.042 | −0.028 | 0.014 | −0.020 | 0.022 |
| (0.006) | (0.004) | (0.009) | (0.003) | (0.008) | |
| Observations | 78 | 78 | 78 | 78 | 78 |
| R-squared | 0.39 | 0.48 | 0.16 | 0.42 | 0.18 |
| Ln SPI | Ln VA | Ln EPI | Ln VA | Ln EPI | |
|---|---|---|---|---|---|
| calibrated using | calibrated using | ||||
| the 10 African countries in Broda, Greenfield, and Weinstein (2017) | Togo and Madagascar | ||||
| (1) | (2) | (3) | (4) | (5) | |
| Ln Remoteness | 0.103 | 0.062 | 0.165 | 0.041 | 0.144 |
| (0.037) | (0.029) | (0.050) | (0.024) | (0.048) | |
| Ln Population | 0.042 | −0.028 | 0.014 | −0.020 | 0.022 |
| (0.006) | (0.004) | (0.009) | (0.003) | (0.008) | |
| Observations | 78 | 78 | 78 | 78 | 78 |
| R-squared | 0.39 | 0.48 | 0.16 | 0.42 | 0.18 |
Notes: The dependent variable is the log spatial price index (Ln SPI) in column (1), the log spatial availability index (Ln VA) in columns (2) and (4), and the log exact price index (Ln EPI=Ln SPI + Ln VA) in columns (3) and (5). Standard errors account for spatial autocorrelation within a 50 km radius around cities using the spatial HAC procedure. , , and respectively, denote significance at the 1 per cent, 5 per cent, and 10 per cent levels.
In large cities, the intensive component (the weighted price index , column (1) of Table 4) is higher, whereas the availability index (the weighted price index , columns (2) and (4) of Table 4) is lower, because more products are available. Which force dominates depends on the value of . Because consumers’ valuation for variety is decreasing in , the intensive price channel dominates for higher values of . For the calibrated values of we use, we find that the price effect dominates the variety effect, which implies that, all else equal, the cost of living increases with city size. The positive relationship between city size and the cost of living is stronger when we use the calibrated elasticities based on the estimates for Togo and Madagascar. This reflects the fact that, as already mentioned, these calibrated elasticities are higher than those obtained when using the ten African countries available in Broda, Greenfield, and Weinstein (2017). In unreported exercises, we find that the two effects exactly cancel out when is set to 3 for all product categories.
Note that if individual prices were not affected by city size (as we found for the barcode products), then the availability effect would unambiguously dominates, and the cost-of-living would always fall with city size.
On the other hand, remote locations exhibit both higher intensive and higher extensive components of the EPI, due to the detrimental effect of remoteness on both the price of available products and product availability. The cost-of-living is thus unambiguously higher in remote locations. The cost of remoteness decreases with though, since the higher the substitution elasticity, the lower the consumers’ love for variety, and thus the less detrimental for consumers’ welfare the unavailable products in remote locations.
Accounting for product availability quantitatively matters. We consider the first calibration of (ten African countries), and we compare two cities at the first and ninth deciles of the distribution in terms of population size and remoteness, respectively. Based on column (1), which only accounts for the price of available products, the estimated coefficients imply that the cost-of-living is 11.0 per cent higher in a large city, and 5.2 per cent higher in a remote one. If we now take the coefficients in column (3) for the exact price index, which accounts for both the price of available products and for product availability, the cost-of-living is 3.7 per cent higher in a large city, and 8.3 per cent higher in a remote one. Hence, once product availability is accounted for, the toll imposed by remoteness on the local cost-of-living is higher than the one related to city size.
5. Conclusion
The strong dispersion in the cost of living across Ethiopian cities we document implies that nominal income should be deflated by a local price index to have a neat view of real income spatial heterogeneity. Importantly, such a local price index should account for product availability. In unreported investigations at the level of Ethiopian regions, we find that deflating nominal income can affect the ranking of regions, and that accounting or not for product availability in the deflator does matter too. Accounting for product availability could also matter for cross-country comparisons of cost-of-living. A more in-depth analysis of these issues is left for further research.
Footnotes
See Ferré, Ferreira, and Lanjouw (2012) and Young (2013) for nominal consumption and nominal income differences across urban and rural areas, and Gollin, Kirchberger, and Lagakos (2021) for spatial differences in terms of non-monetary amenities.
The literature also examines the impact of remoteness on outcomes other than prices or product availability. For instance, Dercon and Hoddinott (2005) show that better access to market towns allows rural households to buy their inputs at a lower price and to sell their outputs at a higher price in Ethiopia; in the same vein, Aggarwal et al. (2022) find poor market access implies a poor harvest output in rural Tanzania.
See also Matsa (2011) on the link between competition and inventory shortfalls.
The correlation between this measure of prices and the average or the mode is above 99%.
We also conducted the Kuiper test to assess deviations of the data from Benford’s Law. The mean Kuiper statistic is 0.0264, which exceeds the threshold above which we can reject with a 99% confidence the null hypothesis that the observed distribution deviates from Benford’s Law.
The travel times to other cities provided by this package are missing for ten cities. For these cities, we estimate travel time based on bilateral distance. Distance is computed using the Stata package Geodist. Travel time is regressed on a polynomial of degree seven of distance, along with fixed effects for origin–destination pairs of regions. For the 8,730 city pairs for which we have both forms of information, the is approximately 90%. Travel time is then predicted for all city pairs involving the ten destinations for which information on travel time is missing. Travel times are those reported by Georoute in 2018.
But not product availability, results available upon request.
A recent literature has examined the role of rationing in developing countries (Gadenne 2020). Rationing involves setting a quota on consumed quantities while imposing a fixed price. This reduces the dispersion across cities in both the availability and prices of products. If anything, this should diminish the coefficients on remoteness and city size. In Ethiopia, less than 15% of cereal production is subject to rationing. When estimating the equations for cereals only, we find that the relationship between prices and remoteness is not significant, but the relationship with city size is. The relationship between product availability and remoteness and population size is similar to that observed for other products. The overall stability of the coefficients, combined with the fact that rationing affects only a fraction of a specific product, suggests that our results are not primarily driven by this policy.
We do not exploit the panel dimension because our measures of remoteness and city size cannot be computed on a yearly basis.
Feenstra, Xu, and Antoniades (2020) show that population size has an unambiguously negative effect on producers’ markups and prices if sunk costs and marginal costs are similar across cities. However, if sunk costs depend on local costs or market size, then prices are less negatively related to market size and could even be higher in larger markets (anti-competitive effect of market size). Unfortunately, with the data at hand, there is nothing we can do to deal with markups. One can also think that the marginal costs inclusive of distribution costs are higher in large cities due to local wages and rents. This could lead to higher prices in large cities, even if producers’ quality and markups remain constant across markets.
The Sato–Vartia weights are given by
Cities for which the index cannot be properly computed have the same median level of remoteness than the rest of the sample, but their population is four time smaller (in median). We thus cannot compute the index for small cities with many missing products, but these cities only account for 8% of the population covered by our sample. Moreover, for the intensive part of the price index (SPI) that can be computed for all cities, we have checked that the coefficients on remoteness and city size are almost the same in the full sample (106 cities) and the constrained sample (78 cities).
Supplementary data
Supplementary data are available at Economic Geography Journal online.
Conflict of interest statement. None declared.
Acknowledgements
We thank Simon Alder, Nate Baum-Snow, Kristian Behrens, Remi Jedwab, François Libois, Marion Mercier, Daniel Pereira Arellano, Tsur Sommerville, Will Strange, Gonzague Vannoorenberghe, and participants at various seminars and conferences for helpful discussions. We also thank the editor and the referees for their insightful comments.
