The World Bank Economic Review, 36(4), 2022, 999–1020 https://doi.org10.1093/wber/lhac018 Article Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Roads and Jobs in Ethiopia Matteo Fiorini and Marco Sanfilippo Abstract Does improving roads affect jobs and structural transformation? A novel geocoded data set covering the uni- verse of Ethiopian roads matched with individual data allows the relationship between improvements in road infrastructure and labor-market outcomes over the 1994–2013 period to be identified. At the district level, greater market access due to better roads correlates with the process of structural transformation in Ethiopia. Improvements in market access are related to reductions in the share of agricultural workers and increases in that of workers in the services sector, but not in manufacturing. Heterogeneity in this relationship exists across industries, gender, education level, and age cohorts. Patterns of internal migration and changes in economic opportunities can help rationalize these findings. JEL classification: L16, O18, O55, R4 Keywords: structural transformation, jobs, roads, Ethiopia 1. Introduction Developing countries can hardly embark on economic development and structural transformation without first reducing transport costs (Gollin and Rogerson 2014). Greater connectivity can improve the lives of individuals, widening their work and educational opportunities, while fostering transition to more productive activities. In addition, improving domestic transport infrastructure can reduce some of the constraints that affect the private sector in many low-income countries, allowing firms to better connect to local and international markets. This, in turn, can improve the efficiency of firms and enable them to offer better jobs. Understanding whether policies supporting the construction of transport infrastructure can affect jobs in low-income countries is therefore a question of high policy relevance. Matteo Fiorini is a programme associate with the Global Governance Programme at the European University Institute, Firenze, Italy; his email address is matteo.fiorini@eui.eu. Marco Sanfilippo (corresponding author) is an associate professor of economics at the University of Torino, Torino, Italy; he is also an affiliate at the Collegio Carlo Alberto, and a programme associate and visiting researcher with the Global Governance Programme at the European University Institute; his email ad- dress is marco.sanfilippo@unito.it. The authors would like to thank the editor, Eric Edmonds, three anonymous reviewers, Margaret McMillan, Ali Sen, Kunal Sen, Victor Stolzenburg, participants of the annual UNU-WIDER Conference at ESCAP in Bangkok, “The Future of Industrial Work” Workshop at UNIDO in Vienna, the Jobs & Development Conference (on- line), the Nordic Conference on Development Economics (online), and seminars at the Universities of Catania and Napoli Parthenope for valuable comments on preliminary draft versions of the paper. They also thank Francesco Iacoella for valuable research assistance. Support from Fitsum Mulugeta and the Ethiopian Economic Association is gratefully acknowledged for facilitating the collection of data from the Central Statistical Agency. Matteo Fiorini acknowledges financial support from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 770680 (RESPECT). A supplementary online appendix is available with this article at The World Bank Economic Review website. © The Author(s) 2022. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 1000 Fiorini and Sanfilippo In this paper, we look at whether and how developments in road infrastructure interact with labor- market outcomes in Ethiopia. We take advantage of the collection of very granular information on a recent large-scale program, the Road Sector Development Programme (RSDP). The RSDP started in 1997, with the aim of improving connectivity across the country through the rehabilitation of existing roads and the Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 construction of new ones.1 In the space of just a decade, the improvements due to the RSDP have been remarkable. Road density rose from 24.1 per 1,000 km2 when the program started to 44.4 in 2010 (when an evaluation of the first three rounds of the program was completed; Ethiopian Road Authority 2011). Over the same period, the proportion of the road network in good condition increased from 22 percent to 56 percent. Our analysis uses geolocalized information on the Ethiopian road network, for which we track specific road-segment improvements undertaken through the RSDP. We match information on the road network with information at the level of individuals. Individual information was taken from the 1994 Population Census and the Ethiopian National Labour (NLF) Survey, a nationally representative survey of Ethiopian workers available for the years 1999, 2005, and 2013. We use the district (or woreda, the third adminis- trative unit level in Ethiopia) as the unit of analysis. To better explore how transport infrastructures affect labor demand, we further combine road data with additional information on the activity of firms. The case of Ethiopia is particularly relevant for our purposes. Beginning with the agricultural- development-led industrialization (ADLI) strategy in 1995, and later with growth and transformation plans, a large emphasis has been attributed to structural transformation. This policy agenda promotes entrepreneurship and diversification into highly productive activities. Improving connectivity both within the country and with external markets occurs in parallel with the pursuit of structural transformations and economic upgrading (Ali 2019).2 Existing evidence from Ethiopia shows that high transport costs have so far represented barriers to market integration (Atkin and Donaldson 2015; Gunning, Krishnan, and Mengistu 2018) and labor supply (Franklin 2018). We study the impact of road infrastructure as an indicator of market access. This allows us to account for the direct and indirect effects of road investments that took place all over the country under the RSDP. Improved access to markets makes locations more attractive to production and consumption, raising population density, the relative price of non-tradable goods (Fajgelbaum and Redding 2018), and, more generally, fostering economic activity (for instance, Storeygard 2016; Alder 2019; Chiovelli, Michalopou- los, and Papaioannou 2019; Eberhard-Ruiz and Moradi 2019). Changes in market access alter the economic environment for both firms and workers, affecting the labor market. Improvements in road infrastructure reduce firms’ transport costs, increasing market oppor- tunities while lowering the cost of sourcing inputs. This can trigger private sector development through increased entry, and higher performance, and ultimately generates an increase in labor demand.3 How- ever, better roads also increase competitive pressures faced by firms, with potentially opposite implica- tions on labor demand. On the supply side, roads can contribute to pushing workers out of agricul- ture, which is still the prevalent source of employment in the country. This happens primarily through improvements in farm productivity (due, for instance, to greater access to new and imported inputs).4 1 Note that roads represent the main transport infrastructure in Ethiopia over the period covered by our analysis. The railway, connecting Addis Ababa to Djibouti, was in fact reestablished in 2017. 2 The recent efforts to develop industrial parks in the country are consistent with the idea that employment creation in non- agricultural sectors and the resulting pattern of structural transformation are indeed dependent on reliable infrastructure. Note, however, that our sample covers a period during which none of the industrial parks were operating. 3 For the case of Ethiopia, Fiorini, Sanfilippo, and Sudaram (2021) show that improvements in market access due to the RSDP are a necessary condition for firms to experience productivity gains from trade liberalization. 4 Recent evidence provides support linking improvements in connectivity under the RSDP with increases in agricultural productivity (Adamopoulus 2020; Gebresilasse 2020). The World Bank Economic Review 1001 In addition, lower transport costs reduce constraints to migration choices (Morten and Oliveira 2018; Lagakos 2020).5 Understanding how labor demand and supply interact in the Ethiopian context in response to improved road infrastructure and the consequences of this on employment is therefore an empirical question that Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 we try to address in the paper. In our empirical analysis we exploit the time-series dimension of improvements in roads within each district. This feature of the data allows us to run a regression with district fixed effects and region-specific time trends, to control for time-contingent shocks and to partial out confounding heterogeneity across districts.6 Many factors can simultaneously concur with improvements in market access, both within and outside a district’s borders. In all of our specifications, we try to minimize endogeneity by controlling for improvements to roads within each district that are orthogonal to changes in the district’s connectivity (similarly to Donaldson and Hornbeck 2016). Still, this does not ensure that our results can be interpreted in a causal way, given that factors unrelated to the characteristics of individual districts can also play a role (e.g., programs targeting remote locations in a process of regional convergence). Hence, in the rest of the paper we are careful to avoid interpreting these as causal effects. Nevertheless, we feel that the relationships are sufficiently interesting and, importantly, policy relevant to justify our analysis. Our results show that some of the changes occurring in the labor market of Ethiopian districts are associated with improvements in road infrastructure. There is no evidence that districts improving market- access experience increases in employment, but we observe changes in the sectoral composition of the workforce. This happens through a reduction of agricultural workers and an increase of workers in the services sector, but not in manufacturing. Improvements in roads seem to go hand in hand with a pattern of structural transformation without manufacturing, which is consistent with the findings of other studies looking at the dynamics of structural transformation in the region (Rodrik 2016; Baccini et al. 2022). At a more disaggregated level, improvements in roads are associated with increases in jobs due to the provision of market services. We also do not find evidence of changes in the composition of jobs due to increases in the construction sector or in government-related activities. In the second part of the analysis, we account for the heterogeneity in individual characteristics. First, we show the existence of gender-specific patterns in our empirical framework. Within the services sector, women seem to respond more quickly to opportunities from improved market access compared to men, a result that confirms previous evidence on the gender-specific benefits of infrastructure (e.g., Dinkelman 2011; Lei, Desai, and Vanneman 2019). Second, we find evidence that larger proportions of the working- age population with a higher education are in districts where investments in road infrastructure provide increased market access, as well as increased participation for the school-age population. Third, disaggre- gating by age cohorts, we show that the main dynamics particularly involve the youngest workers.7 Finally, we investigate some of the potential economic mechanisms that can help to understand our results. By looking at migration patterns, we find evidence of domestic migration to areas characterized by higher market access, along with increases in migrant employment in modern activities. These findings can 5 According to a report by the recently established Ethiopian Jobs Commission (Jobs Creation Commission Ethiopia 2019), increases in migration (mostly rural–urban) do exert a pressure on urban labor markets, with likely consequences on wages, unemployment, and the size of the informal sector. 6 Most of the existing studies on the impact of infrastructures employ a difference-in-differences approach (see, for in- stance, the review by Redding 2022). Dercon et al. (2009), Mu and van de Walle (2011), Faber (2014), and Storeygard and Jedwab (2020) use a specification in first differences that wipes out any fixed effects in the level of the economic outcome of interest. Closer to the empirical approach adopted here, the papers by Alder (2019), Gebresilasse (2020), Aggarwal (2018), and Khandker, Bakht, and Koolwal (2009) instead employ specifications in levels controlling for location and time fixed effects. 7 The main patterns identified in our analysis are robust to several checks, including different specifications, various cuts of the data, and alternative definitions of the variables of interest. 1002 Fiorini and Sanfilippo help explain the structural transformation response to the road infrastructure reform in Ethiopia. Next, using data on both formal and informal manufacturing firms, we find that the response of manufacturing firms to greater market access includes improvements in productivity, as well as a relative increase in the number (and wages) of non-production workers. On the other hand, firms operating in trade-related Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 services (i.e., wholesalers and retailers) respond to higher market access by increasing their average size, but with no change in their productivity. Our work is related to a growing body of research using micro data to investigate the drivers of structural transformation in developing countries (see Lagakos and Shu 2021 for a recent review of the existing evidence). Among those drivers, infrastructure has been widely studied due to its persistent effects on urbanization and the distribution of economic activities across space, as well as due to second-order advantages related to lowering the costs of migration (e.g., Khandker, Bakht, and Koolwal 2009; Adam, Bevan, and Gollin 2018; Bryan and Morten 2019; Hjort and Poulsen 2019; Adukia, Asher, and Novosad 2020; Asher and Novosad 2020; Storeygard and Jedwab 2020).8 Most of all, we contribute to a small strand of evidence looking at the consequences of infrastructural investment in Ethiopia, and in particular under the RSDP. Shiferaw et al. (2015) provide evidence on the positive effects of the RSDP on business dynamism, finding evidence of more entry in the formal sector. Fiorini et al. (2021) show that the reduction in transport costs enabled domestic manufacturing firms to take advantage of trade liberalization, increasing their productivity. Their findings are in line with our mechanisms relative to the manufacturing sector firms. The lack of employment growth in the formal manufacturing sector that we find in our analysis, despite gains in productivity, is consistent with findings by Diao et al. (2021), who attribute it to the diffusion of capital-intensive techniques related to global trends in technology. Adamopoulus (2020) and Gebresilasse (2020) link improvements in connectivity under the RSDP to increases in agricultural productivity using a panel of Ethiopian districts over a similar period to the one we cover.9 Their findings offer a complementary perspective to ours, providing evidence on a mechanism that we cannot test with our data, i.e., increased productivity in agriculture due to reduced transport costs and higher market access.10 The findings of Adamopoulus (2020) are especially relevant for us. He shows that due to a fall in transport costs, production shifts to more productive areas that specialize in the production of cash crops for the export markets, and concentrate in larger farms. This reduces the labor required for food production and generates a structural shift to non-agricultural activities. An important difference with our work is that these two papers look at rural areas and more specifically at the effects of the Universal Rural Road Access Program (URRAP), which was introduced in 2011 with the aim of connecting rural villages to all-weather roads. Much closer to the spirit of the current paper is the work by Moneke (2020), which also looks at the role of infrastructural investments for structural transformation in Ethiopia. There are both differences and similarities in the two studies. First, the scope of his work is broader than ours. His paper investi- gates existing complementarities in road construction and electrification and uses a quantitative model 8 Among these studies, Hjort and Poulsen (2019), Adukia, Asher, and Novosad (2020), and Asher and Novosad (2020) look specifically at how connection to infrastructures can affect structural transformation. Similar to our findings, Hjort and Poulsen (2019) show that this effect is likely driven by the rise of a more dynamic private sector in treated locations. Asher and Novosad (2020) and Adukia, Asher, and Novosad (2020) exploit rich information on the construction of roads in rural villages in India. They offer a more nuanced set of results. Investments in roads do stimulate the realloca- tion of workers out of agriculture and higher investment in education but do not significantly increase local economic activities in treated areas. 9 The paper by Dercon et al. (2009), though not explicitly evaluating the RSDP, shows that access to all-weather roads has important effects in terms of both consumption growth and poverty reduction among Ethiopian rural households. 10 Though they do not test this specifically, there is considerable evidence showing that increases in productivity cause a drop in this sector’s employment and, ultimately, structural transformation (Bustos, Caprettini, and Jacopo 2016). The World Bank Economic Review 1003 to understand their welfare implications. Second, there are differences in how the papers measure road improvements. While Moneke (2020) uses a dummy variable indicating the presence of all-weather roads in a given district, we adopt a market-access approach that allows us to account for both the direct and indirect benefits of road construction and expansion. This allows us to test some specific mechanisms, i.e., Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 migration and firms’ expansion, that are more likely to depend on country-wide, rather than merely local, road improvements.11 Still, despite these differences he also finds that road investments alone promote structural transformation out of agriculture and towards services but not into manufacturing. The remainder of the paper is organized as follows. Data are presented in the next Section, while Section “Empirical Specification” describes the empirical strategy. The key findings of the paper are discussed in the “Results” Section, while Section “Robustness” reports some of the main robustness checks and extensions. Some of the channels potentially driving our findings are discussed in the “Mechanisms” Section. The last Section offers some conclusive remarks. 2. Data Individuals. Individual-level data are obtained by combining two sources that provide complementary information. The first is the Ethiopian National Labour Force (NLF) survey. This is a representative survey of both urban and rural areas administered by the Central Statistical Agency (CSA), with the objective of monitoring the economic and social conditions of the economically active population. The information provided in the survey includes, among other items, the demographic characteristics of the individuals, their education, and working conditions. The NLF includes information on whether respondents report a previous residence different from the current one, thus allowing the identification of internal migrants, as well as on the formal or informal nature of an individual’s current job. We use all existing waves of the NLF, covering the years 1999, 2005, and 2013.12 A limitation of the NLF surveys is that they do not cover the period before the RSDP. To address this issue, we combine the NLF with the 1994 population census, which also provides details on the distribution of workers across industrial sectors.13 Once the NLF surveys and the census data sets were harmonized,14 we collapsed all of the information at the district-year level using sample weights to recover information on the underlying population. Table 1 reports the distribution of labor shares over the 1994–2013 period, computed at the national level using the sample of working-age population. While employment is on the rise, there is also evidence of the process of structural transformation occurring in the country. Over time, workers are less engaged in agriculture and more active in the services sector. However, despite a visible increase in its employment share, the manufacturing sector remains small. The data also show that the relative position of women in the labor market has improved over time (they represent about 47 percent of total workers in 2013, up from 43.8 percent in 1994), especially in the services sector. 11 Moneke (2020) finds no evidence of local roads’ improvements on internal migration, but shows evidence on positive selection of migrants driven by roads. No comparable analysis is performed on firm-level data. 12 The NLF surveys are representative at the national level and use regions, the first administrative units, as the main sample domains. They cover all urban and rural areas of the country except the non-sedentary areas in the Somali region. The sampling frame to select enumerator areas is provided by the population census (the 1994 census for the 1999 and 2005 NLF waves and the 2007 census for the 2013 wave). All of the relevant information on the sampling procedures, coverage, and full descriptive statistics are available in the survey reports published by the CSA (2004, 2006, 2014). 13 This information is not included in the 2007 population census, which we do not use given the specific purposes of our analysis. 14 NLF survey data are not geocoded but include identification codes for each location, including region, zone, and district. To combine the different waves of data, we used the definition of district (woreda) provided by IPUMS that matches districts, using their names when the geographic definition of borders differed between the 1994 and 2007 censuses. Overall, the final estimation sample covers, on average, about 80 percent of the estimated total population in each wave. 1004 Fiorini and Sanfilippo Table 1. Sector Composition of Employment (Percentage) Year Employment Agriculture Manufacturing Construction Services 1994 76.21 87.95 1.99 0.35 9.62 1999 75.69 77.67 4.80 1.05 16.35 Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 2005 81.09 77.89 5.35 1.65 14.99 2013 80.83 70.84 5.17 2.35 21.11 Source: Authors’ elaboration on Ethiopian Central Statistical Agency data. Note: The first column reports the proportion of employed persons in the working-age (15–64) population. Following the Ethiopian National Labour Force report, a worker is defined as a person who declared at least 1 hour of work during the week preceding the interview. The following columns report the proportion of sectoral workers out of the total number of employed persons in the specific year. All data have been weighted before collapsing information at the national level. Figure 1. RSDP Roads in 1996 by Surface Type Source: Authors’ elaboration on Ethiopian Road Authority data. Note: RSDP stands for Road Sector Development Programme. Roads. The main source of information on road infrastructure is a proprietary geospatial database consisting of coded reports by the Ethiopian Road Authority (ERA) covering all road construction and/or rehabilitation sites that were opened under the different phases of the RSDP. The data are organized as a time series of shapefiles of the Ethiopian road network, reporting two main attributes for each geolocalized road segment: the type of road surface and the road’s condition.15 Figure 1 presents the network of federal and regional roads in 1996 by surface type. Figure 2 shows the same types of roads in 2014, distinguishing between segments that existed in 1996 and were not 15 There are four types of road surface in the data: earth surface, minor gravel (which identifies regional rural roads with a gravel surface), major gravel (federal gravel roads), and asphalt. As for road conditions, the database distinguishes between two categories: not rehabilitated and new or rehabilitated. The World Bank Economic Review 1005 Figure 2. New and Upgraded RSDP Roads from 1996 to 2014 Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Authors’ elaboration on Ethiopian Road Authority data. Note: RSDP stands for Road Sector Development Programme. rehabilitated by 2014 (light-grey segments on the map) and roads that were either newly constructed or rehabilitated during the first three phases of the RSDP. A visual inspection of the two maps shows a substantial expansion of the road network between 1996 and 2014. Moreover, road development does not appear to be geographically concentrated but, rather, spans different administrative areas across the country. The information on surface type and condition can be aggregated to compute the average travel speed for each road segment at each point in time. This is done following a standard speed matrix proposed by the ERA and reported in table S1.1.16 We employ an indicator of market access (Donaldson and Hornbeck 2016) to measure the economic effects of infrastructural development in the context of a formal structural gravity trade model. In the context of the present paper, and similarly to Storeygard (2016), market access captures the structure of road connections between a geographically defined area and all other markets in the country, weighted by the intensity of their economic activity. For each district i, market access is defined as the weighted sum of income in each district z different from i, with weights equal to the iz bilateral transport cost scaled down by a trade-elasticity parameter. Formally, Market_Accessit = log D− θ iz,t Lz , z=i where Diz,t is the minimum distance in hours of travel between district i and district z given the road network in place at t. Bilateral distances in travel hours are computed by applying the Dijkstra algorithm 16 The same speed matrix has been used by Shiferaw et al. (2015) and Storeygard and Jedwab (2020). 1006 Fiorini and Sanfilippo on the network of Ethiopian districts (the nodes are set at each district’s centroid) connected by federal and regional Ethiopian roads (links). The variable Lz is an indicator of economic activity based on nightlight intensity in z, and θ is the trade-elasticity coefficient. There are different values of θ in the literature, ranging from 1 to 10 (Donaldson and Hornbeck 2016; Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Chiovelli, Michalopoulos, and Papaioannou 2019) depending on the context. In this paper, we use a trade elasticity of 3.12. We obtain this value following the procedure adopted by Storeygard and Jedwab (2020), i.e., combining the estimated trade elasticity with respect to roads for the United States (1.27) with the difference in cost–travel-time elasticity that has been estimated for Ethiopia by Atkin and Donaldson (2015) (2.46 times the US value).17 In the section on robustness we show, however, that our main results remain consistent with the adoption of different values of θ . While some papers—including Donaldson and Hornbeck (2016)—use population data in the computa- tion of market access, we employ nightlight intensity data as in Storeygard (2016), Chiovelli, Michalopou- los, and Papaioannou (2019), Baum-Snow et al. (2020), and Alder (2019).18 This is particularly appro- priate given that nightlight is a better indicator of local economic development, in contrast to population, which provides improved information on the size of an area but lacks information on its purchasing power (Chiovelli, Michalopoulos, and Papaioannou 2019). With specific reference to Ethiopia, satellite imagery has a better capacity to catch local economic development and population dynamics, especially in lowland areas (about 60 percent of the country’s territory) where part of the population lives in nomadic, semi-nomadic, or pastoral ways, so that official data are less likely to provide precise information.19 Information on nightlight intensity is sourced from NOAA National Geophysical Data Center (2018) and is available at the level of 0.86 km2 grid cells over the whole country area. For each cell the nightlight intensity score can vary from 0 to 63.20 For each district we compute the sum of the nightlight intensity scores across all the cells within the district’s border. We fix the weight Lz at the beginning of the sample period (1994) to exclude potential correlation between changes in destinations’ economic activity and our outcome variables. The mean value of Lz is 6,176 and the median is 3,435, while the minimum and maximum values are 0 and 83,408 respectively.21 A potential drawback of using nightlight instead of population is that the former includes many zeros. Since our unit of analysis—the district—usually includes both urban and rural areas, this is less of a concern compared to more granular settings. Figure 3 plots the value of the market-access indicator at the beginning of our baseline estimation sample (1996) for all Ethiopian districts covered in our estimation sample. Figure 4 shows the change in market access between 1996 and 2014 for each woreda. Focusing on fig. 3, dark blue woredas near the center of the country close to Addis Ababa reveal higher market access in this area. Figure 4 shows a larger increase in market access for less-connected districts away from the center, suggesting that improvements in road infrastructure occurred over the time period of our analysis. By combining information on travel time expressed in hours with data on nightlight intensity, single values of the market-access variable cannot be interpreted in isolation. To get a better sense of the quality 17 In their paper, which looks at the effects of market access on urbanization in Africa, Storeygard and Jedwab (2020) obtain a value of 3.8 because they use the estimated cost–distance elasticity for Nigeria, which is three times larger than that of the United States (1.27*3 = 3.8). Since the paper by Atkin and Donaldson (2015) provides detailed estimates on the cost–distance ratio for Ethiopia (see their table 4), we use that value. 18 Note that other works (e.g., Alder 2019) use market access computed with beginning of sample period weights as an instrumental variable (IV) for the market access using time-varying nightlight density weights. 19 See, for instance, a recent analysis by the World Bank, available at http://devseed.com/ethiopia-docs/. 20 Following Eberhard-Ruiz and Moradi (2019), we use scores from raw satellite images, instead of processed images with stable nightlights, as more reliable proxies of economic activity in small and medium African urban areas. 21 Measuring economic activity via nightlight has been subject to criticism in the literature. For instance, two recent papers (Asher et al. 2021; Gibson et al. 2021) point out that nightlight lacks temporal consistency and is thus less suited to time- series analyses. We note that these concerns are mitigated in our case since we use cross-section variation in nightlight to weight our market-access measure. The World Bank Economic Review 1007 Figure 3. Market Access: Starting Point in 1996 Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Authors’ elaboration on ERA data. Note: The figure plots Market_Accessr,1996 . of road connections behind a certain value of market access, we can look at relevant statistics computed on the distribution of travel times from the origin district and in the year corresponding to that value. For instance, the median year–district observation in terms of market access is 1996 Ganta Afeshum in the northern Tigray region, with a value of 6.01968. The average travel time from Ganta Afeshum to all other districts in our sample in 1996 was 39.5 hours, with a minimum of 3.5 hours. The observation with the highest value of market access in our sample is 2014 Arada (with a value of 10.946), followed by 2014 Lideta (10.904). Arada and Lideta are two districts in the region of Addis Ababa. The average travel time to connect them with all other districts in our sample in 2014 was approximately the same for both and is equal to 13.7 hours, with minima of less than 20 minutes. Finally, the observation with the lowest value for market access is 1996 Moyale (3.163). Moyale is a district in the Somali Region. This woreda includes the southernmost point of the whole country, on the border with Kenya. The average travel time from Moyale to all other districts in our sample in 1996 was 55 hours, with a minimum of 19 hours. 3. Empirical Specification The objective of our empirical analysis is to study the link between improvements in connectivity, captured by variation in market access, and district-level labor-market outcomes in Ethiopia. For each outcome variable, we propose the following baseline specification: yit = β Market_Accessit + γi + ρrt + it , (1) where y captures a generic outcome variable among those included in our panel of Ethiopian dis- tricts i across years t. These include the proportion of the population employed and the percentage of 1008 Fiorini and Sanfilippo Figure 4. Market Access: Change by Woreda (Market_Accessr,1996 ) Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Authors’ elaboration on ERA data. Note: The figure plots the difference Market_Accessr,2014 − Market_Accessr,1996 . employment in agriculture, manufacturing and services. The term Market_Accessit is the measure of con- nectivity between district i and relevant economic activity in the rest of the country at time t. Each spec- ification includes district fixed effects and region-specific time trends. District fixed effects are important to control for all the time-invariant characteristics of the district (e.g., geophysical features, such as soil quality and elevation) that can simultaneously affect the decision to invest in roads and labor-market outcomes. Region-specific time trends account for common changes (e.g., regional policies, or changes in regional budget on roads) that can confound the relationships among the outcomes and the treatment. In estimating equation (1), standard errors are clustered at the regional level.22 Due to the small num- ber of regions (n = 11), all estimation tables in the paper report wild cluster bootstrap standard errors (Cameron, Gelbach, and Miller 2008),23 although we show in the section on robustness that the results are consistent with different clustering strategies. Finally, all regressions are weighted by each district’s population. Our estimation sample consists of an unbalanced panel of 1,573 observations covering 506 districts. Taken together, these observations account for over 80 percent of the total population and total jobs in the country. Table S1.2 in the supplementary online appendix reports descriptive statistics of our outcomes of interest and of the main regressors. 22 In this we follow Abadie et al. (2017), and we cluster according to the sampling strategy of the NLF surveys, which are representative at the regional level. 23 We implement this in STATA using the boottest routine written by Roodman et al. (2019) The World Bank Economic Review 1009 A potential threat to identification in our empirical setting is the endogeneity of the main regressor of interest, Market_Accessit . While the fixed effects capture the main sources of omitted-variable bias, potential confounding heterogeneity at the district–time level remains an active source of endogeneity in our specification. Reverse causality can also play a role, with time-contingent shocks to local employment Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 and/or economic activity shaping incentives for investment in local roads. For instance, geographic areas with relatively larger (smaller) agricultural or services sectors might be systematically more (less) interested and successful in attracting infrastructure investment for the improvement of local roads in the district. We follow an identification strategy similar to the one proposed by Donaldson and Hornbeck (2016). More precisely, we exploit the fact that variation in each district’s market access is determined by improve- ments to the whole national road network (as we keep market size fixed, market access does not respond to changes in economic activity over time). Moreover, when captured in the market-access measure, im- provements in local road segments reflect not only the higher road coverage and/or traveling speed but also their contribution to the district connectivity relevant to economic activity in the rest of the country. We can therefore partial out the changes in the quality of the local road network, which are a major source of endogeneity concerns. Indeed, one might hypothesize that political economy forces at the local level lead to both changes in aggregate and/or sectoral employment within the district and to investments in local roads. We capture the district-level infrastructure developments through a weighted sum of the distance covered by each road segment within the district area, with weights equal to the speed allowed by the type of surface and the road condition. We denote this variable by Local_Roads.24 While it is fair to assume a positive relationship between market access and local roads,25 the linkages between local roads and indicators of structural transformation are not trivial. When fixing a district’s connectivity with respect to relevant economic activity in the rest of the country, it is not clear whether economic forces within the district, activated by improvements of the local road network, are sufficient to trigger labor reallocation.26 On the other hand, after controlling for local roads, the coefficient of the Market_Access variable identifies the relationship between within-district measures of structural trans- formation and any change in the Ethiopian road network that (a) affects district-level connectivity with relevant economic activity in the country and (b) is in principle orthogonal to the mere expansion of the road network within the district. 4. Results 4.1. Jobs and Structural Transformation We start by introducing a set of results linking market access to the number and sectoral composition of jobs in Ethiopian districts. The dependent variable yit is, in turn, the share of total jobs in the working-age population in district i at time t and the share of jobs in each of the main sectors of the economy over the 24 Our strategy based on partialing out follows Donaldson and Hornbeck (2016). However, its rationale is also fully consistent with other works (see, for instance, Storeygard and Jedwab 2020) that build a measure of market access without factoring in improvements in local roads and use this measure as an instrument for overall market access. The literature on the effects of transport infrastructure has advanced other solutions to address endogeneity concerns, including identification strategies relying on time-invariant instruments, such as historical or planned infrastructural networks (see Redding and Turner 2015; Redding 2022, for a review). However, those strategies are more likely to capture variation in the location of infrastructure, rather than the evolution of investment over time (Storeygard and Jedwab 2020). 25 The empirical correlation between market access and local roads is positive and statistically significant. In a simple univariate linear regression of market access on local roads, the estimated coefficient is equal to 0.076 with a robust standard error of 0.004. When the same relationship is estimated including district fixed effects and region-specific time trends as in (1), we get an estimate of 0.022 with a standard error of 0.004. 26 Existing work looking at the role of local roads has mostly done so at the level of individual towns or villages (e.g., Asher and Novosad 2020). 1010 Fiorini and Sanfilippo Table 2. Roads and Jobs Jobs Agriculture Manufacturing Services Dependent var. (1) (2) (3) (4) (5) (6) (7) (8) Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Market_Access 0.008 0.012 −0.046 −0.049 0.007 0.008 0.033 0.035 (0.012) (0.013) (0.006) (0.004) (0.003) (0.003) (0.004) (0.005) Local_Roads — −0.002 — 0.001 — −0.000 — −0.001 (0.001) (0.002) (0.000) (0.001) Wild p-value Market_Access 0.578 0.438 0.0117 0.0127 0.348 0.238 0.0107 0.00293 Wild p-value Local_Roads — 0.203 — 0.496 — 0.297 — 0.604 Mean DV 0.808 0.808 0.790 0.790 0.0409 0.0409 0.154 0.154 Quantification 0.00828 0.0132 −0.0496 −0.0532 0.00722 0.00858 0.0356 0.0382 Observations 1,573 Source: Authors’ elaboration on data from the Ethiopian Central Statistical Agency and the Ethiopian Road Authority. Note: The dependent variables measure, respectively, the ratio between the number of jobs and the working-age population in each district (Jobs), the share of agricultural workers in total workers (Agriculture), the share of manufacturing workers in total workers (Manufacturing), and the share of service workers in total workers (Services). All regressions include district fixed effects and region-specific time trends. All regressions are also weighted by the size of the district population. Standard errors clustered at the region level are reported in parentheses below each estimated coefficient. Wild p-values indicate the p-value for wild cluster bootstrap standard errors at the region level. Mean DV is the sample mean of the dependent variable. Quantification reports the change in the dependent variable associated with an increase in Market_Access of 1 sample standard deviation. number of total jobs. For each dependent variable, table 2 reports estimates from two specifications. The first is the baseline regression featuring market access on the right-hand side, in addition to district and region–year fixed effects (equation (1)). The second is the same specification augmented with a measure of road segment improvements within the district (Local_Roads). Standard errors clustered at the region level are reported in parentheses below each estimated coefficient while p-values based on wild cluster bootstrap standard errors are reported at the end of the table. There is a small difference in both the size and the precision of the estimated coefficients for market access when comparing the two models. This suggests that what matters in the relation between market access and the outcomes of interest is not district-specific changes in the length and speed allowed on local roads but changes at the country-level road network that increase the district’s connectivity to economic activity in the rest of the country. As for local roads, the wild p-values reported in columns 2, 4, 6, and 8 in table 2 suggest that improvements in the local road network that are orthogonal to changes in the district’s connectivity have little implication for structural transformation.27 The lack of statistical significance for the estimated coefficients reported in columns (1) and (2) of table 2 suggest that job creation is not correlated with within-district improvement in market access over time. On the other hand, we find some evidence of correlations between market access and structural transformation. Indeed, there is evidence of lower shares of agricultural workers in districts reporting increases in market access. The decrease in the share of agricultural jobs seems to occur in relation to an increase in the services sector, rather than manufacturing.28 This pattern is not uncommon in low-income 27 The results presented in table 2 remain robust when removing the population weights, and working on a balanced sample of districts (i.e., excluding those changing borders or denomination over the sample period). The estimates derived in these robustness tests are not reported for reasons of space, but are available upon request. 28 Working with the shares of sectoral employment does not allow us to draw an unequivocal conclusion about the nature of the reallocation we observe in the data. Two additional tests that we have run can help to understand the process. First, we have replicated the analysis using absolute values (i.e., using the total number of jobs by district). The results of this exercise (not included but available upon request) are not conclusive either. The direction of all sectoral coefficients is consistent with the main findings, but only the manufacturing sector displays a significant (positive) value. Second, when disaggregating by age cohorts, we show that most of the changes we observe can be explained by young people entering the job market. The latter finding is discussed in more detail in the next subsection. The World Bank Economic Review 1011 Table 3. Unpacking the Services Sector Private Public Utilities Construction Others Dependent var. (1) (2) (3) (4) (5) Market_Access 0.013 0.000 0.001 0.005 0.010 Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.001) (0.002) (0.001) (0.004) (0.006) Local_Roads −0.000 0.001 0.000 −0.000 −0.001 (0.000) (0.000) (0.000) (0.000) (0.001) Wild p-value Market_Access 0.0312 0.848 0.299 0.213 0.467 Wild p-value Local_Roads 0.359 0.0801 0.368 0.576 0.463 Mean DV 0.0585 0.0265 0.00231 0.0131 0.0233 Observations 1,573 Source: Authors’ elaboration on data from the Ethiopian Central Statistical Agency and the Ethiopian Road Authority. Note: The dependent variables measure, respectively, the share of workers employed in private services (Private), in public services (Public), in the utilities sector (Utilities), in construction (Construction), and in other services (Others). Private services include the following industries: trade, financial, real estate, and transport. Public services include the following industries: public administration, education, and health. Others is a residual category including personal services. All regressions include district fixed effects and region-specific time trends. All regressions are also weighted by the district population size. Standard errors clustered at the region level are reported in parentheses below each estimated coefficient. Wild p-values indicate the p-value for wild cluster bootstrap standard errors at the region level. Mean DV is the sample mean of the dependent variable. countries and echoes existing evidence on the direction of structural change, which shows a reallocation of workers out of agriculture towards services. The estimated coefficients are sizeable. According to the estimates in columns (4) and (8), an improve- ment of market access by 1 sample standard deviation correlates with a reduction of about 5 percent- age points (p.p.) in the share of agricultural workers and an increase of about 4 p.p. in the share of workers employed in the services sector. These numbers are economically significant as they represent 7 percent and 25 percent of the sample average shares of agricultural and service workers, respectively. Note that a 1 standard deviation change in Market Access corresponds to moving from the sample me- dian of market access to its 80th percentile, i.e., from the starting sample value of Ganta Afeshum in the Tigray region and almost at the border with Eritrea, to the end sample value of Kalo, a district of the Amhara region in the middle of the motorway connecting Addis Ababa to Mekele (the capital of Tigray). 4.1.1. Services To unpack the baseline results presented above, we further ask which specific services sectors are most affected by roads. Table 3 provides a summary of regression estimates after grouping relevant subsets of two-digit sectors. We find that improvements in market access mainly correlate with higher shares of workers in private services, a group of services including trade-related (wholesale and retail) activities, financial and business services. Conversely, we do not find evidence of road-driven improvements in indus- tries that can be directly connected to road construction. This is also relevant for identification purposes. It shows that our results are not mechanically driven by an increase in jobs in the construction sector itself or in related services. An important issue in our context is the fact that investment in infrastructure is often accompanied by additional public services (e.g., maintenance, security, provision of utilities) that can create new jobs directly linked to the infrastructure investment. Provided that these services belong to the public sector, the statistically insignificant coefficient in column 2 of table 3 shows that public services are not driving the baseline results in our sample. This confirms that the pattern of structural change captured in our results reflects improvements in access to markets rather than the increase of non-market services provided by the government. 1012 Fiorini and Sanfilippo 4.2. Heterogeneity Gender. We test for potential heterogeneity in our main results when distinguishing individuals by gender. The estimates reported in table S1.3 show that the patterns described in the previous section are largely confirmed across genders. While men are more likely to leave agricultural jobs and join both services and Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 manufacturing, the coefficient of services denotes a larger propensity for women to be driven into mod- ern activities following changes in market access. The coefficient of market access is statistically different across the two specifications for the gender-specific share of services jobs.29 Evidence from developing countries shows that women face greater difficulties in the labor market compared to men and are dis- proportionately affected by infrastructural bottlenecks. Improving connectivity can reduce some of these constraints, saving time spent in unpaid activities and enabling opportunities beyond the local community (Lei, Desai, and Vanneman 2019). Education. Increases in market access can shape educational investment decisions. They may increase the number of people returning to education on the one hand, while increasing the opportunity cost of schooling on the other (Adukia, Asher, and Novosad 2020). In our empirical framework, we explore whether and how changes in market access affect educational choices. Educational indicators cover infor- mation on the highest grade completed and the current grade attended. The number of individuals with some level of education has been growing over time. However, only a very small fraction of individuals report an education level higher than primary school (grades 1 to 8). We run two different exercises with education data. The first measures changes in the share of employed individuals with different levels of education. The estimates are presented in table S1.4 and show that participation in the labor market by better-educated workers is significantly increasing in areas with improving market access over the sample period.30 The second exercise analyzes specific cohorts of individuals, namely those that are of school age, i.e., from 7 to 18.31 By carrying out this analysis, we find some evidence of a positive correlation between higher levels of education for children in districts experiencing greater market access (see table S1.5). Age cohorts. Related to the previous exercise, another important source of heterogeneity is the de- mographic composition of the working-age population. Since a large share of the working population is young, it is important to understand whether economic opportunities are most likely to involve young workers or not. Hence, we spilt our sample and replicate our baseline analysis for the following three age cohorts: 15–19, 20–39, and 40–65.32 Results, reported in table S1.6, show that some of the change is indeed occurring in the younger cohorts of workers. The youngest cohort, in particular, is also likely to experience increases in their employment rate (though starting from lower levels) in relation to an increase in market access. 29 To test whether the estimated coefficients are statistically different across gender-specific specifications, we have ap- pended the data for female- and male-specific versions of each dependent variable. Then we have estimated our baseline model augmented with (a) a gender indicator for the dependent variable and (b) an interaction with that indicator and any other term on the right-hand side of the model. The coefficient for the interaction between the gender indicator and market access is statistically different from 0 (p-value equal to 0.032) in the specification for the share of services sectors in total jobs. 30 Though our data do not allow further inferences, these results suggest a higher return to education in areas with better connections and are consistent with the recent work by Adukia, Asher, and Novosad (2020) linking investment in roads to educational outcomes in rural India. 31 Since 1994, Ethiopia has had an 8-2-2 formal education structure. Primary school has an official entry age of 7 and a duration of eight grades. Secondary school is divided into two cycles: lower and upper, which consist of grades 9– 12. Students sit the Primary School Certificate Examination at the end of grade 8, the General Secondary Education Certificate Examination at the end of grade 10, and the Higher Education Entrance Certificate Examination at the end of grade 12 (the source of this information is UNESCO). 32 The three age groups represent, respectively, 26.3 percent, 50.1 percent, and 23.6 percent of the population in our sample. The World Bank Economic Review 1013 5. Robustness Growth hubs. As noted by Faber (2014) and Storeygard and Jedwab (2020), growth in economic hubs might drive the location of road placement and implementation. Hence, we run some robustness checks in which we remove potential growth hubs from the sample. We first exclude the woredas in the Addis Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Ababa special administrative division (six in total). Next we also exclude the districts where regional capitals are located.33 Finally, we replicate our estimates excluding all districts belonging to the Tigray region, which hosts the majority of the Tigrayan ethnic group that was in political power until 2018. This test is motivated by the political economy argument, according to which co-ethnicity can drive public investment choices (Burgess et al. 2015). Results, summarized in table S1.7, show that our main findings hold across all of these altered data sets. Alternative measure of market access. We explore whether our results are robust to an alternative mea- sure of market access. Specifically, we experiment with different values of trade elasticity (θ ). Following the existing empirical evidence, we use three alternative values of θ : (a) a value of 1, as originally proposed by Harris (Harris 1954); (b) a value of 1.5, which was adopted by Gebresilasse (2020) in his work mea- suring the effects of market access under the RSDP and the URRAP programs in rural Ethiopia; (c) a value of 8.22, which was used by Donaldson and Hornbeck (2016) in their work on railroads in the United States. Finally, we also report results based on a definition of market access that replaces nightlights with population as an indicator of economic activity in destination markets.34 The results are reported in table S1.8. While the magnitude and precision of the estimated coefficients change, the results do not provide significant evidence that contradicts the qualitative pattern suggested by our baseline analysis. Alternative definition of local roads. In the spirit of Donaldson and Hornbeck (2016), controlling for changes to local roads allows us to isolate the variation in market access that is orthogonal to investments in each district. However, our main definition of local roads—given by the total (speed-weighted) length of all roads within the borders of the district—may not capture investments developed in nearby areas outside the district’s administrative borders that still reflect district-specific incentives or are undertaken in expectation of a district-specific payoff. A recent paper by Storeygard and Jedwab (2020) argues that choosing closer rather than farther buffers will introduce a trade-off in terms of excludability versus strength in the identification strategy. To address this issue, we augment our baseline specification by controlling for improvements in all roads, as captured by the same variable computed considering the district area extended by buffers of 10 to 50 km. We replicate our baseline estimates by adding the resulting controls in the same regression including our main measures of market access and local roads. Table S1.9 reports one of these estimates, including a buffer of 50 km, which largely confirms the baseline patterns discussed above. Different clustering of the standard errors. In this section we check whether the results survive different clustering strategies of the standard errors. These include clustering standard errors at the district level (i.e., the level of treatment) or using heteroskedastic robust methods. In addition, we check whether the possible presence of spatial correlation in the residuals can affect the results. To do this, we estimate our model by introducing a spatial HAC correction of standard errors based on the Conley method, using the code proposed by Hsiang, Meng, and Cane (2011). We impose no constraints on the temporal decay of the weights and test the robustness of our specification to different lengths of the radius (respectively, from 100 to 500 km) for the spatial kernel. Table S1.10 reports the results. For the Conley method, we 33 The capital of the Oromia region was moved to Addis Ababa in 2005. Still, for the purpose of this exercise we use the old capital, Adema. As in the case of Addis Ababa, we also include all districts (two in total) belonging to the special administrative zone of Dire Dawa. 34 Data on population is taken from 1994 census. 1014 Fiorini and Sanfilippo only report results based on a 150 km cutoff, but standard errors are generally smaller when considering greater distances (especially in the specifications for services) and our results remain qualitatively stable.35 Omitted variables and reverse causation. Although the nature of our results is mostly descriptive, a potential issue of concern is the bias in the estimated coefficient that can occur. This may be due to either Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 the omission of time-invariant variables at the district level affecting the relationship between roads and labor-market outcomes and/or the reverse causation among the two. To address the former issue, we run a specification that includes several controls. One is nightlight intensity, which is a commonly adopted proxy for the level of economic activity at the subnational level. We also account for the number of conflicts occurring in each district on a yearly basis. Finally, we include information on weather conditions, namely the level of yearly precipitation (in millimeters) and the average monthly air temperature (in degrees Celsius).36 Results, reported in table S1.11, show that the overall findings remain largely unaffected by the inclusion of these controls. We then run some exercises to understand to what extent the presence of pre-trends, or the influence of initial conditions, may affect our findings. First, we estimate the relationship between the overall sample change in our treatment and baseline values of the outcome variables. More precisely, we run the following regression: Market_Accessi = β Xi + φi + i , where the dependent variable is the change in market access from 1996 (i.e., one year before the beginning of the RSDP) to 2013 and Xi is a vector of initial characteristics of district i. These include our main outcome variables, i.e., initial employment and the shares of agriculture, manufacturing, and services in total jobs. Initial characteristics are computed using information included in the 1994 census. We also run an alternative version of the previous equation in which the delta of market access is computed comparing each period t with the previous, and regress this value on employment and sectoral shares measured in t − 1. Estimates of both exercises are reported in panels A and B respectively of table S1.12 and show no evidence that the main outcomes of interest are driving subsequent investment in road infrastructure. Second, to better deal with pre-trends, we run an exercise in which interaction terms between time trends and initial values of the outcome variables are included as additional regressors. This should help alleviate the concern that districts with, for instance, high initial agricultural employment prior to the RSDP may experience differential structural transformation trajectories over the period of the program.37 Results, reported in table S1.13, show that the inclusion of initial values interacted with time dummies does not alter the size or the direction of the findings. 6. Mechanisms Our main results provide a detailed characterization of the role of infrastructure reforms in the process of structural transformation in Ethiopia. In particular, we find a correlation between improvements in market access and an employment transition from agriculture towards services (especially for women). We also find some evidence of higher investment in education in more intensively treated locations. In this section we try to harmonize these pieces of evidence by testing some of the potential underpinning mechanisms. There are several channels through which improvements in market access can affect changes 35 Results of the additional specifications based on different cutoffs are not included due to space considerations but are available upon request to the authors. 36 Data on conflicts and weather is provided by Aiddata GeoQuery. Conflict data is originally sourced from ACLED Conflict Events; data on precipitation is from the UDel Precipitation data set (v.5.01); data on temperature is from the UDel Air Temperature data set (v.5.01). 37 This is a commonly used strategy in papers examining the implications of trade liberalization (e.g., Dai, Huang, and Zhang 2021). The World Bank Economic Review 1015 in the composition of the labor force: directly, by lowering transport costs and therefore reducing some of the typical frictions affecting labor mobility and internal migration in developing countries, and indirectly, by leveling the playing field through increased economic opportunities and competition in the treated locations. Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 6.1. Internal Migration Migration, especially from rural to urban areas, is related to changes in economic development and nor- mally fuels the processes of structural transformation and urbanization in developing countries (Gollin and Rogerson 2014; Storeygard and Jedwab 2020). However, high transportation costs and a lack of economic opportunities might hamper migration (Lagakos 2020). Hence, we ask whether improvements in market access, by driving down transportation costs and increasing economic opportunities at destina- tions, can facilitate the movement of workers towards areas with better economic prospects. While this is intuitive, there is very little empirical evidence supporting this relationship. An exception is the recent work by Morten and Oliveira (2018) showing how improved market access had a large role in abating migration costs in Brazil. We investigate this potential channel in two ways. First, we check whether improvements in roads are conducive to more migration in treated districts. Information available in both the population census and the NLF data allows us to track past changes in the respondents’ places of residence. An individual is classified as an (internal) migrant in the year of the survey if their birthplace is different from the place where they currently reside. For this exercise, we can only track migrants at the destination and not at the origin, as information on the former is only available at a more aggregated geographic area (the zone). While migration could be motivated by several reasons, the search for work opportunities is—according to the qualitative information provided by the NLF surveys—the main one, and it was also on the rise over the period examined (see table S1.14). Results in table S1.15 seem to confirm that locations with higher levels of market access are likely to attract a larger number of migrants, irrespective of whether they work or not. In columns (2) and (3), we split migration according to whether their location within a given district is urban or rural. The results are statistically significant only for urban migration, confirming that improvements in roads are more likely to make urban locations more attractive. Second, we replicate our main specification replacing total workers with migrant workers. This shows whether changes in the labor-market outcomes that are correlated with improvements in market access reflect the increase in the relative share of migrant workers. Results are reported in table 4 and show that migrant workers are (a) more likely to be employed following improvements in market access and (b) more likely to be engaged in non-agricultural activities, especially in manufacturing and services. 6.2. Economic Opportunities We now turn to the channel that explains the role of market access for structural transformation through its ability to shape economic opportunities and labor demand. To frame this analysis, we match the road data with firm-level data sets covering manufacturing sectors and trade-related services.38 6.2.1. Manufacturing Firms Data on manufacturing firms come from two sources. The first is the annual census of large and medium manufacturing establishments, published by the CSA. Manufacturing industries are defined at the four- digit level according to the ISIC Rev. 3 classification. Data cover all formal firms with at least 10 persons employed and that use electricity in their production process.39 These firms are required to respond to this census every year; therefore, it reports on all large and medium firms in the manufacturing sector. 38 Table S1.16 reports the summary statistics on all the variables used in the firm-level analysis. 39 The number of persons refers to employees as well as working owners. 1016 Fiorini and Sanfilippo Table 4. Migrant Workers Employment Agriculture Manufacturing Construction Services Dependent var. (1) (2) (3) (4) (5) Market_Access 0.015 −0.004 0.003 0.001 0.015 Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.009) (0.004) (0.002) (0.002) (0.005) Local_Roads 0.001 0.002 −0.000 −0.000 −0.000 (0.002) (0.001) (0.000) (0.000) (0.001) Wild p-value Market_Access 0.00781 0.455 0.00391 0.676 0.0371 Wild p-value Local_Roads 0.398 0.0107 0.453 0.424 0.961 Mean DV 0.128 0.0751 0.00875 0.00340 0.0445 Observations 1,573 Source: Authors’ elaboration on data from the Ethiopian Central Statistical Agency and the Ethiopian Road Authority. Note: The dependent variables measure, respectively, the share of migrants in the total work force (Employment), the share of migrants working in the agricultural sector (Agriculture), the share of migrants working in the manufacturing sector (Manufacturing), the share of migrants working in the construction sector (Construction), and the share of migrants working in the services sector (Services). All regressions include district fixed effects and region-specific time trends. All regressions are weighted by the size of the district population. Standard errors not corrected for the wild cluster procedure are reported in parentheses below each coefficient. The p-values for wild cluster bootstrap standard errors at the region level are reported at the end of the table. Table 5. Share of Informal Manufacturing Sector Year % of firms % of employment % of value added 2006 97.14 51.42 38.77 2008 96.74 59.64 31.06 Source: Authors’ elaboration on Ethiopian Central Statistical Agency (CSA) data. Note: All values represent the share of informal manufacturing firms on the total values. The latter is given by the sum of informal and formal firms’ annual totals. Information on informal firms is calculated using sample weights provided by the CSA. We report the information only for the two years in which the survey of small-scale manufacturing industries and the firm census were run simultaneously. Value added is computed as the total value of production minus production costs. The census records provide information on the characteristics of each establishment, as well as detailed information on the size and composition (including by skills and gender) of the workforce and on the location of each firm.40 Our data cover yearly information over the 1998–2009 period, ending a few years earlier than the analysis conducted so far. The second data set is the survey of small-scale manufacturing industries (SSIS). We combine all existing waves of the SSIS, covering the years 2002, 2006, 2008, 2010, and 2014. This is a survey that covers small (i.e., those employing less than 10 persons) and informal firms in the manufacturing sector. The sample is single-stage stratified, considering six main industries (textiles and garments, metalwork, woodwork, leather and leather products, other manufacturing sector, and grain mills industry), sampled in similar proportions across regions. Due to the lack of a proper sample frame, it is not necessarily representative of the sector but provides considerable information on the activities of smaller firms, which comprise the majority of firms in the country. Over 95 percent of the firms surveyed in the different waves of the SSIS do not keep a book of accounts (or declare it incomplete) and are hence informal. On average and consistently over time, small and informal firms represent the large majority of all manufacturing establishments, approximately half of total manufacturing employment, and about one-third of the value added produced in the sector. Table 5 reports precise figures for the two years in which the SSIS and the census were run simultaneously. Information included in the SSIS is based on a similar questionnaire to the census of larger firms, allowing for the comparison of some outcomes. 40 Note that information on the location of manufacturing firms is slightly more precise than that used in this paper, so we have computed market access at the level of the (urban) town, rather than at the district level. The World Bank Economic Review 1017 We start with the census data by showing whether improvements in market access at firm locations are related to the dynamism of the private sector and several dimensions of firm performance. In all regressions we control for firm and region–year fixed effects as well as for a measure of local roads. Results are presented in table S1.17. In the first two columns, we collapse the data at the location–year level and Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 show that while improving road infrastructure is not correlated with firm entry, it correlates with the entry of foreign firms. It has been shown that foreign investors normally generate higher-quality jobs, pay higher salaries, and create more links with local firms (including in the services sector).41 Moving to the firm-specific changes, those in locations that have improved market access are found to have experienced gains in a few dimensions, including an increase in (labor) productivity. Moreover, we find some (though statistically weak) evidence of a compositional shift towards non-production workers, whose real per capita wages are also positively correlated with increases in market access. Next we replicate the previous set of exercises using data on informal manufacturing firms. This is particularly relevant since informal firms account for a large majority of all manufacturing firms, as well as for a (slight) majority of the sector’s employment (see table 5). Due to the lack of a panel dimension in the SSIS data, we run our analysis using district, industry (at the four-digit level of the ISIC classification) and region–year fixed effects to understand how aggregate and average firm indicators have changed over time within the same district under differential changes in market access. The results, reported in table S1.18, are to some extent similar to those obtained by looking at formal firms. These findings suggest that improvements in market access are related to improvements in informal firm productivity. On the other hand, there is no clear evidence regarding the composition and wages of workers. 6.2.2. Services Firms No equivalent information is available for the services sectors. Our analysis of services firms is therefore based on the Ethiopian Distributive Trade Survey (DST), which covers firms in trade-related services (i.e., retailers and wholesalers) and is available in a cross-sectional setting for the years 2003, 2009, and 2011. Moreover, this survey only covers urban areas and, therefore, is not representative at the national level.42 However, it includes information on the districts in which firms are located, as well as other basic information about their activity, such as size, sales, capital, and wages of employees. Based on these data, we conduct a similar analysis to the one discussed for manufacturing firms. Esti- mates presented in table S1.19 show that services firms in districts experiencing improvements in market access do experience increases in their average size. There is no evidence of corresponding changes in wages or labor productivity. This suggests an increase in employment in private services and it is there- fore consistent with the findings reported in table 3. 7. Conclusions In this paper we have studied the relation between road infrastructure development on the size and com- position of jobs in Ethiopian districts. We have taken advantage of novel geocoded information covering the Ethiopian road network, which we combined with individual information from population censuses and nationally representative labor-force surveys. Our analysis has generated the following results. Higher market access at the district level due to road development is related to a process of structural transfor- mation characterized by a reduction in the share of agricultural workers in favor of services. We do not find evidence of improvements in market access being correlated to more jobs in the manufacturing sec- tor. We also show that such changes are most likely to benefit women and younger individuals, and that 41 Looking at the local impact of Foreign Direct Investment (FDI) in Ethiopia, recent work by Abebe, McMillan, and Serafinelli (2022) provides sound evidence that the entry of FDI generates high spillovers on domestic firms and workers. 42 Urban areas covered in the survey correspond to 15 major urban centers (regional capitals and other major towns) and 106 towns. 1018 Fiorini and Sanfilippo better road infrastructure brings about a potential upgrading of the labor force through higher partici- pation in education. We highlight our results by showing that higher economic activity induced by road investments stimulates both the demand from firms, through increases in their size and productivity, and supply from workers, who are more likely to migrate towards areas in the country with greater market Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 access. Overall, our results show that investments in road infrastructure under the RSDP can support the process of job creation and structural transformation since they can contribute to reducing some of the typical frictions affecting the labor market in Ethiopia. Yet the weak role of roads in raising the manu- facturing sector’s capacity to generate jobs is concerning, especially in view of the country’s high political focus on industrialization. However, it is important to highlight that our analysis does not cover the most recent years, when most of the industrial parks (e.g., the one established in Hawasa) have started large-scale activities and for which it is well known that infrastructure investments are key. Further re- search is needed to investigate the effects of infrastructure in more recent periods and to gain a better understanding of the relationship between market access and structural transformation in low-income countries. Data Availability Statement Access to individual and firm data is provided under the signing of a NDA with the Central Statistical Agency of Ethiopia. Raw data on the road sector are proprietary. For replication, we are able to share all the final measures and all the codes to construct them, along with the full replication files for all tables and figures in the paper. References Abadie, A., S. Athey, G. W. Imbens, and J. Wooldridge. 2017. “When Should You Adjust Standard Errors for Cluster- ing?” NBER Working Paper N. 24003. Abebe, G., M. McMillan, and M. Serafinelli. 2022. “Foreign Direct Investment and Knowledge Diffusion in Poor Locations: Evidence from Ethiopia.” Journal of Development Economics, https://doi.org/10.1016/j.jdeveco. 2022.102926. Adam, C., D. Bevan, and D. Gollin. 2018. “Rural–Urban Linkages, Public Investment and Transport Costs: The Case of Tanzania.” World Development 109: 497–510. Adamopoulus, T. 2020. “Spatial Integration and Agricultural Productivity: Quantifying the Impact of New Roads.” mimeo. Adukia, A., S. Asher, and P. Novosad. 2020. “Educational Investment Responses to Economic Opportunity: Evidence from Indian Road Construction.” American Economic Journal: Applied 12 (1): 348–76. Aggarwal, S. 2018. “Do Rural Roads Create Pathways out of Poverty? Evidence from India.” Journal of Development Economics 133: 375–95. Alder, S. 2019. “Chinese Roads in India: The Effect of Transport Infrastructure on Economic Development.” mimeo. Ali, S. N. 2019. “Infrastructure and Economic Transformation in Ethiopia.” In The Oxford Handbook of the Ethiopian Economy. Oxford University Press. Asher, S., T. Lunt, R. Matsuura, and P. Novosad. 2021. “Development Research at High Geographic Resolution: An Analysis of Night-Lights, Firms, and Poverty in India Using the Shrug Open Data Platform.” World Bank Economic Review 35 (4): 845–71. Asher, S., and P. Novosad. 2020. “Rural Roads and Local Economic Development.” American Economic Review 110 (3): 797–823. Atkin, D., and D. Donaldson. 2015. “Who’s Getting Globalized? The Size and Implications of Intra-National Trade Costs.” NBER Working Paper N. 21439. Baccini, L., M. Fiorini, B. Hoekman, and M. Sanfilippo. 2022. “Services and Economic Development in Africa.” The World Bank Research Observer, https://doi.org/10.1093/wbro/lkac006. The World Bank Economic Review 1019 Baum-Snow, N., V. Henderson, M. A. Turner, Q. Zhang, and L. Brandt. 2020. “Does Investment in National Highways Help or Hurt Hinterland City Growth?” Journal of Urban Economics 115: 1–19. Bryan, G., and M. Morten. 2019. “The Aggregate Productivity Effects of Internal Migration: Evidence from Indone- sia.” Journal of Political Economy 127 (5): 2229–2268. Burgess, R., R. Jedwab, E. Miguel, A. Morjaria, and G. P. i Miquel. 2015. “The Value of Democracy: Evidence from Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Road Building in Kenya.” American Economic Review 105 (5): 1817–51. Bustos, P., B. Caprettini, and P. Jacopo. 2016. “Agricultural Productivity and Structural Transformation. Evidence from Brazil.” American Economic Review 106 (6): 1320–65. Cameron, A. C., J. B. Gelbach, and D. L. Miller. 2008. “Bootstrap-Based Improvements for Inference with Clustered Errors.” Review of Economics and Statistics 90 (3): 414–27. Chiovelli, G., S. Michalopoulos, and E. Papaioannou. 2019. “Landmines and Spatial Development.” mimeo. Colella, F., R. Lalive, S. O. Sakalli, and M. Thoening. 2019. “Inference with Arbitrary Clustering.” IZA Discussion Paper n. 12584. CSA. 2004. “Analytical Report of the National Labour Force Survey.” Statistical Bulletin of the Central Statistical Authority. ——— . 2006. “Analytical Report of the National Labour Force Survey.” Statistical Bulletin of the Central Statistical Authority. ——— . 2014. “Analytical Report of the National Labour Force Survey.” Statistical Bulletin of the Central Statistical Authority. Dai, M., W. Huang, and Y. Zhang. 2021. “How Do Households Adjust to Tariff Liberalization? Evidence from China’s WTO Accession.” Journal of Development Economics 150: 1–21. Dercon, S., D. O. Gilligan, J. Hoddinott, and T. Wodehanna. 2009. “The Impact of Agricultural Extension and Roads on Poverty and Consumption Growth in Fifteen Ethiopian Villages.” American Journal of Agricultural Economics 91 (4): 1007–21. Diao, X., M. Ellis, M. McMillan, and D. Rodrik. 2021. “Africa’s Manufacturing Puzzle: Evidence from Tanzanian and Ethiopian Firms.” HKS Working Paper No. RWP21-018. Dinkelman, T. 2011, December. “The Effects of Rural Electrification on Employment: New Evidence from South Africa.” American Economic Review 101 (7): 3078–108. Donaldson, D., and R. Hornbeck. 2016, 02. “Railroads and American Economic Growth: A “Market Access” Ap- proach.” Quarterly Journal of Economics 131 (2): 799–858. Eberhard-Ruiz, A., and A. Moradi. 2019. “Regional Market Integration and City Growth in East Africa: Local but No Regional Effects?” Journal of Development Economics 140: 255–68. Ethiopian Road Authority. 2011. “Road Sector Development Program 13 Years Performance and Phase IV.” ERA. Faber, B. 2014. “Trade Integration, Market Size, and Industrialization: Evidence from China’s National Trunk High- way System.” Review of Economic Studies 81 (3): 1046–70. Fajgelbaum, P., and S. J. Redding. 2018. “Trade, Structural Transformation and Development: Evidence from Ar- gentina 1869-1914.” NBER Working Paper, 20217. Fiorini, M., M. Sanfilippo, and A. Sudaram. 2021. “Trade Liberalization, Roads, and Firm Productivity.” Journal of Development Economics 153: 1–18. Franklin, S. 2018. “Location, Search Costs and Youth Unemployment: Experimental Evidence from Transport Subsi- dies.” Economic Journal 128 (614): 2353–79. Gebresilasse, M. 2020. “Rural Roads, Agricultural Extension, and Productivity.” mimeo. Gibson, J., S. Olivia, G. Boe-Gibson, and C. Li. 2021. “Which Night Lights Data Should We Use in Economics, and Where?” Journal of Development Economics 149: 1–12. Gollin, D., and R. Rogerson. 2014. “Productivity, Transport Costs and Subsistence Agriculture.” Journal of Develop- ment Economics 107 (2014): 38–48. Gunning, J. W., P. Krishnan, and A. T. Mengistu. 2018. “Fading Choice: Transport Costs and Variety in Consumer Goods.” CEPR Press Discussion Paper No. 12910. Harris, C. D. 1954. “The, Market as a Factor in the Localization of Industry in the United States.” Annals of the Association of American Geographers 44 (4): 315–48. Hjort, J., and J. Poulsen. 2019. “The Arrival of Fast Internet and Employment in Africa.” American Economic Review 109 (3): 1032–79. 1020 Fiorini and Sanfilippo Hsiang, S., K. Meng, and M. Cane. 2011. “Civil Conflicts Are Associated with the Global Climate.” Nature 476: 438–41. Jobs Creation Commission Ethiopia. 2019. “State of Jobs 2019.” Federal Democratic Republic of Ethiopia, Addis Ababa. Khandker, S. R., Z. Bakht, and G. B. Koolwal. 2009. “The Poverty Impact of Rural Roads: Evidence from Bangladesh.” Downloaded from https://academic.oup.com/wber/article/36/4/999/6711587 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Economic Development and Cultural Change 57 (4): 685–722. Lagakos, D. 2020. “Urban-Rural Gaps in the Developing World: Does Internal Migration Offer Opportunities?” Journal of Economic Perspectives 34 (3): 174–92. Lagakos, D., and M. Shu. 2021. “The Role of Micro Data in Understanding Structural Transformation.” STEG Pathfin- ing Paper. Lei, L., S. Desai, and R. Vanneman. 2019. “Women’s Non-agricultural Employment in India: The Conditioning Role of Community Gender Context.” Feminist Economics 25 (4): 94–125. Moneke, N. 2020. “Can Big Push Infrastructure Unlock Development? Evidence from Ethiopia.” mimeo. Morten, M., and J. Oliveira. 2018. “The Effects of Roads on Trade and Migration: Evidence from a Planned Capital City.” NBER Working Paper N. 22158. Mu, R., and D. van de Walle. 2011. “Rural Roads and Local Market Development in Vietnam.” Journal of Develop- ment Studies 47 (5): 709–34. NOAA National Geophysical Data Center. 2018. “Nighttime Lights Time Series.” Available at https://ngdc.noaa.gov/ eog/dmsp/downloadV4composites.html. Redding, S. J. 2022. “Trade and Geography.” In Handbook of International Economics, 5:147–217. Routledge. Redding, S. J., and M. A. Turner. 2015. “Transportation Costs and the Spatial Organization of Economic Activity.” In Handbook of Regional and Urban Economics, 5:1339–98. Elsevier. Rodrik, D. 2016. “Premature Deindustrialization.” Journal of Economic Growth 21 (1): 1–33. Roodman, D., M. Ø. Nielsen, J. G. MacKinnon, and M. D. Webb. 2019. “Fast and Wild: Bootstrap Inference in Stata Using Boottest.” Stata Journal 19 (1): 4–60. Shiferaw, A., M. Söderbom, E. Siba, and G. Alemu. 2015. “Road Infrastructure and Enterprise Dynamics in Ethiopia.” Journal of Development Studies 51 (11): 1541–58. Storeygard, A. 2016. “Farther On Down the Road: Transport Costs, Trade and Urban Growth in Sub-Saharan Africa.” Review of Economic Studies 83 (3): 1263–95. Storeygard, A., and R. Jedwab. 2020. “The Average and Heterogeneous Effects of Transportation Investments: Evi- dence from Sub-Saharan Africa 1960-2010.” NBER Working Paper N. 27670.