The World Bank Economic Review, 39(1), 2025, 104–123 https://doi.org10.1093/wber/lhae017 Article Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 Better Roads, Better Off? Evidence on Upgrading Roads in Tanzania Christelle Dumas and Ximena Játiva ABSTRACT Spatial isolation is considered to be one of the main determinants of poverty. Therefore, many transport in- vestments are undertaken with the stated objective of poverty reduction. This paper evaluates the effect of a Tanzanian program that rehabilitated 2,500 km of major roads on rural livelihoods. The analysis uses a large set of variables describing household behavior in order to provide a complete picture of the adjustments. The identification consists of combining a household fixed effects strategy with propensity score matching. Some damaging effects of the program are found on the rural population in the two years following the intervention: the price of rice decreases; households reallocate labor away from agriculture and provide more wage work, but the increase in wage income does not compensate for the loss in agricultural income. Nor do households seem to be benefiting from the fall in the price of rice at consumption level. These results are consistent with rural households facing increased competition due to reduced transportation costs. JEL classification: O13, O18, J43, O12, O15 Keywords: roads, poverty, rural households, Africa 1. Introduction Eighty percent of the world’s extreme poor and 75 percent of the moderately poor live in rural areas (Castañeda et al. 2018). Therefore, isolation is considered one of the main contributors to poverty. Ac- cording to the World Bank’s Rural Access Index, only 34 percent of the rural Sub-Saharan African popula- tion live within 2 km of an all-season road. Spatial isolation can impose serious constraints on agricultural production and access to health, education, and work opportunities. As a consequence, infrastructure in- vestment has long been considered a key aspect of development policy.1 In this paper, we ask whether Christelle Dumas (corresponding author) is a professor at University of Fribourg, Switzerland; her email address is christelle.dumas@unifr.ch. Ximena Játiva is a researcher at UNICEF Innocenti –Global Office of Research and Foresight; her email address is xjativa@unicef.org. The authors thank Stefan Bauernschuster, Lorenzo Casaburi, Denis Cogneau, Martin Huber, Sebastian Krautheim, Julien Labonne, Sylvie Lambert, Karen Macours, Ferdinand Rauch, Stéphane Straub, Dominique van de Walle, Oliver Vanden Eynde, and Joachim De Weerdt for helpful comments and suggestions. They also thank seminar participants at the University of Fribourg, Paris School of Economics, LICOS, Passau, Heidelberg, Cergy-Pontoise, Graduate Institute and participants of the following conferences: Swiss Development Economics Network, German Development and Economics Policy (Berlin), EEA-ESEM (Manchester), ESPE (Bath), Nordic Development Economics (Copenhagen), DIAL (Paris), NOVAfrica (Lisbon), NCID (Madrid), and EUDN (Toulouse). A supplementary online appendix is available with this article at The World Bank Economic Review website. 1 The World Bank Transport Business Strategy for 2008–2012 states that “one of the best ways to promote rural devel- opment is to ensure good accessibility to growing and competitive urban markets” (The World Bank 2008). C The Author(s) 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. The World Bank Economic Review 105 upgrading a road benefits the local population. Compared with building an entirely new infrastructure, upgrading an old road may only marginally reduce transport costs. However, roads require complemen- tary services such as transportation means and a network of buyers and sellers to allow for an efficient exchange of goods. This paper will therefore provide evidence on the effect of accessing a road of higher quality when these services are already provided. Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 A growing literature seeks to identify the consequences of transportation infrastructure, be it railroads, highways, or rural roads. Most of the literature concludes that better infrastructure has positive effects. Better connectivity to other locations has been shown to lower the prices and increase the availability of non-local goods (Aggarwal 2018; Khandker, Bakht, and Koolwal 2009), increase the use of inputs such as fertilizers and hybrid seeds (Shamdasani 2021; Shrestha 2020), increase agricultural production (Khandker, Bakht, and Koolwal 2009), facilitate agricultural specialization and market orientation of farm activity (Qin and Zhang 2016; Shamdasani 2021), and increase income (Donaldson 2018; Qin and Zhang 2016). Firms located close to major roads are more likely to export (Volpe Martincus, Carballo, and Cusolito 2017; Storeygard 2016), and manufacturing activity is stimulated (Ghani, Goswami, and Kerr 2016). Only a few articles find no or very limited effects (Asher and Novosad 2020; Banerjee, Duflo, and Qian 2020). However, isolation also limits competition from external producers and therefore might allow the subsistence of non-competitive rural farmers. The improvement of transport infrastructure could well have damaging impacts on small-scale farm producers, at least in the short run when sectoral reallocation has not yet taken place. In this paper, we evaluate the effects of a rehabilitation program that upgraded 2,500 km of major roads in Tanzania on agricultural decisions, market participation for labor and products, prices, migration, consumption, and household well-being. We are not aware of any articles identifying the effect of road rehabilitation programs rather than construction programs. Our paper is also a valuable addition to the existing roads literature for the following reasons. First, the effect of transportation on development has largely been assessed in Asia. The scarcity of the literature on Sub-Saharan Africa is all the more striking since the transportation sector accounted for 18 percent of World Bank lending in 2017. Only a few papers address the effect of roads on African rural livelihoods. Casaburi, Glennerster, and Suri (2013) assess a rural road upgrading program in Sierra Leone and show that it reduced market prices of local crops, while Nakamura, Bundervoet, and Nuru (2020) and Dercon et al. (2009) conclude that a rural road program increased consumption in Ethiopia. Fiorini and Sanfilippo (2022) show that better roads reallocate workers from the agricultural sector to the service sector.2 Second, the existing literature on Africa has mostly focused on rural roads, rather than major roads. Our paper addresses the effect of major roads on rural household livelihoods. To our knowledge, the only paper that has the same objective is Shrestha (2020) on Nepal, who finds that reducing the distance to the highway increases land values and market participation for rural households. We do not necessarily expect similar consequences for the improvement of rural roads and major roads. For instance, rural roads could well improve access to other villages and therefore improve the spatial integration of the labor market, while a major road would fail to have such consequences simply because the major road remains too distant to modify labor-market opportunities. The literature has globally emphasized that the last-mile connection to the village is the crucial element (Minten, Koru, and Stifel 2013). On the other hand, spatial integration of the goods market may be wider with major roads and such investment could then trigger a larger reduction in the price of goods than rural roads. The strength of the various mechanisms surely depends on the context and calls for an assessment in Africa. 2 Burgess et al. (2015) also show that road placement is driven by ethnic favoritism in Kenya, Jedwab and Storeygard (2021) find that cities grow more when they have better market access, and Fiorini, Sanfilippo, and Sundaram (2021) show that input tariff reduction in Ethiopia has a greater impact on manufacturing firms’ productivity where better roads improve market access. 106 Dumas and Játiva One difficulty in the road literature consists in addressing the endogeneity of road placement. Roads are likely not placed randomly, and evaluating their impact requires some understanding of the allocation decision. Specifically, there are two types of endogeneity issues. The first is the broad allocation decision on which area to serve with a (better) road. The second is the exact route of the road. In this paper, we do not need to address the latter since the program consists in upgrading roads for which the route Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 is fixed. The endogeneity threat therefore comes from the choice of which road to improve. A series of papers has exploited regression discontinuities in the allocation of roads in India (Asher and Novosad 2020; Aggarwal 2018; Chaurey and Le 2021; Casaburi, Glennerster, and Suri 2013). However, this can only be implemented in a setting where there are a large number of roads to be built and has only been used for rural roads. Regarding major roads or highways, recent papers instrument the exact route by a straight line between main cities or the cost-effective route between cities (Chandra and Thompson 2000; Banerjee, Duflo, and Qian 2020; Bird and Straub 2020; Shrestha 2020), often using historical data to identify nodal points. In this paper we use household panel data and estimate the impact of road rehabilitation conditional on household fixed effects. To increase the likelihood that the within-household change would be similar between treated and non-treated households in the absence of the treatment, we build a comparison group for the treated households using households who also live close to a major road and weight them according to propensity score matching. We show that the road selection criteria match the objectives that were stated by the government agency. We thus compare the evolution of households that live at a given distance from an upgraded road to similar households that live close to a road that was not upgraded. While the matching procedure is satisfactory for balancing household characteristics, some outcome variables remain unbalanced between the treatment and control groups. In particular, the rehabilitation program seems to have taken place in rice-producing areas; this limits our ability to interpret our estimates in a causal way. Our results suggest that rural households, instead of reaching new markets, actually face a decrease in the price of rice, reduce their farm activity, and increase non-farm work. We also find suggestive evidence that this is associated with a decrease in agricultural income, partly compensated by an increase in wage income. We do not find any pattern of migration due to the program. We also provide a series of tests to assess the validity of our identification strategy. We test the common trends assumption in several data sets; the general finding is that it is usually not rejected but some variables do display different trends. We also run placebo tests with roads that were identified as potentially rehabilitated but were not improved over the same time frame and find no effect. This article is structured as follows. The next section provides a conceptual framework; section Context describes the agricultural context of our study and the road improvements in Tanzania during the period 2008–2013. Then we present the data. The empirical strategy is discussed, before we provide the results. The last section discusses the results and concludes the paper. 2. Conceptual Framework We start by offering a conceptual framework that may help to place the results in perspective. Improving roads is expected to reduce transportation costs and therefore transaction costs in exchanges. We focus on the effects on the labor market and the goods market and expect other markets (land, credit, insurance) to be only marginally impacted by such reductions in transaction costs. Asher and Novosad (2020) sum- marize the expected effects of improvements in rural roads by noting that (a) prices and wages should move toward prices outside the village, (b) the expectation is that wages will rise, prices will increase for exported goods and decrease for imported goods and (c) many additional economic effects may take place, potentially reversing this broad expectation (e.g., due to general equilibrium effects at the village level). A few differences between the Tanzanian program and the context in Asher and Novosad (2020) are worth considering. First, the reduction in transaction costs will differ substantially in the labor and goods The World Bank Economic Review 107 markets for a rural road and for a major road improvement. A priori, the change in the labor market will remain limited with a major road improvement. Qualitative fieldwork in rural areas of Bukoba3 reveals that farm labor markets are local, and no laborer provides wage work outside the boundaries of his/her village.4 The improvement of a major road is therefore unlikely to meaningfully impact households who do not live by the road in terms of labor opportunities. By comparison, the effect of the major road is Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 expected to be stronger for the goods market. A grocer is usually available in villages, selling goods that are not produced locally, and middlemen visit villages to buy crops that are produced. These services do not require commuting between the main road and the village on a daily basis, and lower transaction costs on the major road may be reflected in village prices. The relative weight between the effects in the two markets is a key difference between the two settings. Second, the analysis of agricultural households in the presence of market imperfections (Singh, Squire, and Strauss 1986; de Janvry, Fafchamps, and Sadoulet 1991) suggests that a reduction in transaction costs would increase the use of markets and in general allow households to reach a higher utility level by approaching the optimal allocations of labor, consumption, and production. For the goods market, a large part of production is self-consumed, and with lower transaction costs, households will sell more if the market price compensates for the lost value of own-consumption. They may need to reallocate their consumption to exploit the decrease in prices. Changes in price levels of produced goods will necessitate adjustments on the production side to exploit their comparative advantage. If sectoral reallocation were high, then households could easily cope with such changes in their economic environment, but a host of evidence shows that reallocation is particularly slow in developing countries.5 In practice, in the short to medium term, some rural households may suffer from increased price competition instead of benefitting from greater market access. This will be the case if they use a low technology, which prevents them from being competitive. While in the longer run, this process may foster growth through creative destruction, it may be extremely costly for vulnerable segments of the rural population. Given that labor markets are local, we expect wages to reflect the competitiveness of the village in the broad economy and not necessarily to increase when transportation costs decrease. Finally, we anticipate differences between improving a road and building one. A priori, a new road should lead to a greater reduction in transaction costs, particularly for households that were previously far from a road. An improvement in the quality of a road will have only a limited effect, by comparison, since all households were already served by a road. However, roads in themselves do not reduce costs. Costs can only be reduced when a network of buyers and sellers enables goods to be exchanged efficiently. It can be assumed that this network is already in place when the road is built, and that it would come into play immediately when the road is improved. It is therefore not clear whether, in the very short term, a new road would have a greater effect than the improvement of an existing road. In the longer term, we expect the effect of a new road to be greater than that of an upgrade. 3. Context 3.1. Agriculture in Rural Tanzania Over two-thirds of Tanzanians live in rural areas where agriculture serves as their main livelihood activ- ity and accounts for 70 percent of their income. Livestock and poultry production jointly represent the second-largest source of income. Agriculture is the cornerstone of the Tanzanian economy; it employs 74 3 Bukoba is located in the northeast of Tanzania and is the capital of the Kagera region. The fieldwork took place in May 2022. 4 Wage work in services and industries might take place outside the village when people are employed on a permanent basis. 5 See, for instance, Revenga (1997), Harrison and Hanson (1999), Attanasio, Goldberg, and Pavcnik (2004), Currie and Harrison (1997), Topalova (2010). 108 Dumas and Játiva percent of the country’s labor force and accounts for 31 percent of GDP and 20 percent of total annual export earnings. The vast majority of agricultural output (75 percent) is produced by small family farmers, producing mostly for home consumption and selling the excess in local markets. Tanzanian smallholders sell on average only 35 percent of their agricultural production. The main food crops are maize, cassava, rice, sorghum, and bananas. Local and informal markets are the main selling and buying channel for Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 smallholders for both agricultural inputs (99 percent) and outputs (98 percent) (Government of Tanzania 2011; FAO 2018, 2020). Table S2.1 (in the supplementary online appendix) provides descriptive statistics on the agricultural production and consumption of rural households in Tanzania. This table is obtained from our data set, described below. We see that maize is the most important crop, both in value and in kilograms. A large share, but not all of the production, is consumed. The second crop is rice, and the consumption is of the same order of magnitude as production. All the other crops, even cash crops such as cotton and groundnuts, generate comparatively less value than cereals. Given the focus of this paper on rice and maize production, we describe in greater detail the specificities of these two markets. According to Dercon (1998), rice has a higher return than maize but is also riskier. This good is more appreciated by households but also more expensive, as a kilo of rice costs double that of maize. McCullough et al. (2022) find that rice in Tanzania is a luxury good while maize is a necessity good.6 Rice is more commercialized than other staple food crops. In 2002–2003, 42 percent of rice production was marketed against 28 percent of maize and 18 percent of sorghum (Minot 2010). Rice prices in local markets are estimated to be connected to international prices, while maize prices are more driven by local conditions (Minot 2010; Baffes, Kshirsagar, and Mitchell 2017). Overall, Tanzanian maize and Tanzanian rice are competitive compared to neighboring countries (figs S2.1 and S2.2), but Tanzanian rice is often more expensive than Pakistani rice. However, the Tanzanian government interferes with international trade in these crops in many ways. Between 2008 and 2012, three export bans were imposed for maize exports, at times when maize prices were high (Baffes, Kshirsagar, and Mitchell 2017). Tanzania usually imports rice from Pakistan, but since tariffs are extremely high (75 percent), a large share of imports are illegal and transit through neighboring countries. Official figures for internationally traded volumes of maize and rice are thus vastly underestimated, as is clear from the comparison with figures from neigh- boring countries (SERA 2012, 2016).7 In the period under study, the maize price has globally increased, leading to difficulties for households in financing their main staple. However, the increase in total food basket expenditures was contained (Rudolf 2019). 3.2. The Road Upgrading Program The road upgrading program was implemented in a context where infrastructure is considered extremely poor: Tanzania ranked 118th out of 134 economies in the infrastructure dimension of the World Economic Forum’s Global Competitiveness Index in 2008–2009 (Schwab and Porter 2008). This was due to both a low road density (96.5 m/km2 compared to 296.95 in Kenya and Uganda) and to the poor condition of the road network: only 36.63 percent of the road network was paved or sealed8 and was classified as in “good or fair condition” (Government of Tanzania 2008). The 2005 National Strategy for Growth and the Reduction of Poverty in Tanzania identified the poor condition of the network as one of the major impediments to development. Therefore, the plan emphasized that investing in adequate road infrastructure had the potential to boost the rural economy, promote growth, and reduce the level of rural poverty. In this light, the government launched the 10 Year Transport 6 The income elasticity of maize is 0.57 against 1.46 for rice. 7 Exports of rice to Tanzania reported by all exporting countries are two to three times as large as imports reported by Tanzania. Maize figures are even more disconnected from the counterpart estimations. 8 A sealed road has a surface treated with asphalt concrete, chipseal, tarmac, or bitumen. The World Bank Economic Review 109 Sector Investment Programme, which identified nine road transport corridors (approximately 10,300 km of roads) that were key to enhancing the integration of regions within the country, as well as to better connecting the country to its neighbors. During the 10 years of the program, all roads in the nine corridors that were not in good condition (6,000 km) were planned to be upgraded to paved standards. The first phase was implemented between 2007/2008 and 2012/2013 under the responsibility of the Tanzanian Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 National Roads Agency (TANROADS). The guidelines to prioritize the road projects were the following: first, fully funded projects related to the maintenance of roads in poor condition were given priority. Second, road segments that would enhance exports and key food-crop production had to be given preference. Third, all regional centers should be linked with paved roads, and district headquarters should be connected with all-weather roads of at least gravel standards. Finally, the projects should open up all major population areas to the modern economy and trade. We will seek to proxy for these guidelines with a set of georeferenced variables.9 In practice, for each section of road to be upgraded or rehabilitated, experts were commissioned to an- alyze the road’s current surface and structure, the expected traffic flow, and the environmental conditions. All these parameters were taken into account when comparing the various improvement options, with the objective of maximizing cost-effectiveness.10 These improvements could therefore involve improving embankments, improving drainage, adding new lanes, changing the road surface, and/or widening lanes. In addition, the country’s current policy is to prioritize the maintenance of all roads in good condition. This policy is based on the fact that, in the past, neglecting to repair simple failures led to complete dete- rioration of the road, which eventually required capital investment for reconstruction. Our prior is thus that the current policy guarantees that the quality of upgraded roads remains high for at least a few years. Figure S2.3 shows that by the end of 2013, the network showed significant improvement: 2,564 km of roads were paved. Figure 1 shows the state of the network before (panel 1a) and after (panel 1b) the infrastructure work. Prior to the upgrading program, the majority of the network was either graveled (49.7 percent) or sealed (30.5 percent). Only 11 percent of the network was paved and in good condition, whereas at the end of 2013, this percentage had risen to 29.4. Tanzania moved from 118th to 90th place on the infrastructure dimension of the Global Competitiveness Index in 2016–2017 (Schwab and Sala-i Martin 2016). 4. Data We use two main sources of data: the geolocated information on roads obtained from TANROADS and a household panel survey (LSMS-ISA). 4.1. TANROADS Data The Tanzanian Roads Agency collected a georeferenced data set on road condition and upgrades. It covers the entire road network and contains primary and secondary roads, as well as information about their length, surface condition (paved, sealed, gravel, and soil type), upgrading projects, and date of completion. For the purpose of this study, we focus on primary/major roads for which upgrading started and was completed between 2008 and 2013. 9 A second assessment of infrastructure needs and works was scheduled to start in 2012–2013 and end in 2016–2017. However, a review of recent data on public works shows that it has actually been postponed. This Source: Tanzania National Roads Agency, http://tanroads.go.tz. We do not have the same degree of detail for these road upgrades and therefore cannot use them in our analysis. 10 An example of such an analysis is available in the following report: https://www.afdb.org/fileadmin/uploads/ afdb/Documents/Project- and- Operations/Tanzania_- _Singida- Babati- Minjingu_Road_Upgrading_Project_- _Appraisal_Report.PDF 110 Dumas and Játiva Figure 1. Road Rehabilitation 2008–2013. Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 Source: Authors, based on TANROADS data. Note: Figure (a) shows the road network in 2008 and Figure (b) the road network in 2013. . The World Bank Economic Review 111 Table 1. Outcome Variables in the LSMS, by Recall Period Labor market Product market Welfare Household Annual information Annual information Annual information Off-farm labor Harvest (kg, TSh)a Migration Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 Hired-in labor Area (acres)a Household on-farm labor Sales (kg, TSh)a Household farm income Last 7 days Last 7 days Last 7 days Household wage labor Consumption (kg)a Subjective welfare Household income Durable assets Productive assets Community Last 7 days Prices Source: Authors’ own elaboration. Note:This table lists the variables used as outcomes in the analysis, by recall period and by unit of observation.a These variables are computed by crop: rice, maize. 4.2. LSMS-ISA Data The LSMS-ISA (Living Standards Measurement Study – Integrated Surveys on Agriculture) data for Tan- zania consist of a panel of three rounds of a nationally representative household sample collected by the Tanzanian National Bureau of Statistics. The survey samples 3,265 households in the first wave, clustered in 410 enumeration areas across mainland Tanzania; 258 of these enumeration areas are rural and in- cluded in the analysis. The three first rounds were collected in 2008–2009, 2010–2011, and 2012–2013. The attrition rate is low in the panel, as only 4.84 percent of households surveyed in 2008 were not observed in 2010 or 2012.11 These data contain highly detailed information on the household, agricultural, and community dimen- sions, and they are geolocated. Table 1 summarizes the outcome variables by recall period. Regarding agricultural activities, we use the information on the previous long rainy season (the period from Jan- uary to August). The survey collects all data related to inputs and outputs for each household plot. We aggregate this information at the household-year level. Information pertaining to market labor supply and household consumption is collected for a shorter reference period (typically for the week before the survey). Prices for food and standard goods are collected at the village level at the time of the survey. All variables used in the study are described in the supplementary online appendix S1. 4.3. Additional Data We collected a set of proxies for the determinants of the upgrade allocation. Table S2.2 lists the variables and their corresponding sources: we georeferenced major cities, district headquarters, and major border crossings and obtained the 2010 population density and IFPRI standardized agroecological zones, both at 100 m × 100 m resolution. We also build variables such as average labor market activity and education level at the village level from the Labour Force Survey (LFS) collected in 2006. Finally, the data were also matched with rainfall data12 to control for productivity shocks. We checked that the treatment does not correlate with shocks. As there is no such correlation and the estimates are the same with and without controlling for shocks, we simply provide the estimates without this additional control. 11 The dates of collection were Oct 2008–Sept 2009, Oct 2010–Sept 2011, and Oct 2012–Nov 2013. The last round, collected in 2014–2015, refreshes the sample and retains only one-third of the original sample. 12 Rainfall estimates are given by squares of roughly 10 km × 10 km. These data were obtained from http://iridl. ldeo.columbia.edu/SOURCES/.NOAA/.NCEP/.CPC/.FEWS/.Africa/.DAILY/.ARC2/.daily/.est_prcp/datafiles.html. The shock is defined as the normalized deviation from the mean over the period 2001–2013. 112 Dumas and Játiva 5. Estimation Strategy In this section we describe in detail the identification strategy. In short, we run a difference-in-difference estimation with time and household fixed effects, weighted with propensity score to create baseline balance between treated and untreated locations. Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 5.1. Defining the Treatment: Timing In 2008, the program had not yet started, and we therefore have a pretreatment period. A small number of roads were upgraded before the second round of the panel, and considerably more roads were upgraded before the third round of the panel. However, the duration of the survey for each round exceeded one year, and therefore the actual treat- ment depends on the exact date of interview of the household. The treatment is defined as follows: for outcomes that are measured on a yearly basis, the household is treated if the road was completed at least one year before the interview; for outcomes that are measured on a short-recall period, the household is treated if the road was completed at least one month before the interview. Figure S2.4 shows the time line of the program, the periods of data collection, and the long rainy seasons (LRS). Some additional roads were upgraded after 2014, but we only use them in the placebo tests. It is clear from the figure that no household would be considered treated in the 2010–2011 LSMS round for yearly outcomes and that only a limited number of them may be considered treated for short-recall period outcomes. We discard these observations (36 households). 5.2. Estimation The treatment variable defines who is and who is not treated when a road is upgraded. There is prob- ably a continuous effect depending on the distance of the household to the road. However, households living farther away from existing roads might substantially differ from those who are closer. For this rea- son, we implement three complementary strategies. First, we select control households among those who live at a similar distance from a (non-upgraded) road as the treated households. Second, we control for household fixed effects to limit bias arising from remaining latent differences between treated and non- treated households. The strategy is therefore extremely close to a difference-in-difference strategy, except that we control for any unobserved household characteristics that are not time varying. This does not guarantee that the evolution of these households would have been similar in the absence of an upgrade. Therefore, third, we implement a matching strategy to increase the comparability of treated and control households. The identifying assumption is that this control group would have had a similar evolution to the treated group in the absence of the program (Heckman and Navarro-Lozano 2004; Smith and Todd 2005; Caliendo and Kopeinig 2008). Implementing matching requires having a binary treatment and the use of a distance threshold to define the treatment. However, we can perform the exercise for many thresholds and do so for all consecutive kilometers between 20 km and 50 km. We cannot use a lower threshold because the estimation would rely on too few treated households. Because the number of treated households is lower for annual outcomes, we can only provide an assessment of the road program for treatments starting at 30 km and going up to 50 km. Figure S2.5 plots the survey dates in the last round, along with the share of treated households (using 30 km as the cutoff) in each month. Let us use the 261 households surveyed in May 2013 as an example: 76 percent of them live less than 30 km away from a road; 4 percent are considered treated for annual outcomes, while 23 percent are considered treated for short-recall variables. This implies that 19 percent of the sample is considered treated for short-recall variables but serves as control for annual variables. While this design might be surprising, this is because annual variables provide information on a different period, namely the long rainy season of 2012 (January to August). One limitation of our work is that we The World Bank Economic Review 113 interpret our results as if they were obtained on the same sample, which is not the case. The main analysis uses the as-the-crow-flies distance to define the treatment, as it does not require additional assumptions on travel time along secondary roads. Below, we also provide the results using travel time to the main road to define the treatment. Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 5.2.1. Specification Our specification is the following: Yivt = α + δ1d 1it (t = 2012)Roadivtd + δ0d 1it (t = 2010)Roadivtd +θ Zivt + φi + βt + iv t , (1) where Yivt is an outcome variable for household i in village v at time t. The outcome variables were listed previously . The variable Roadivtd is a dummy indicating whether the household is located within a radius of d km of an upgraded road, with d ranging from 20 to 50.13 The term β t includes round fixed effects as well as survey-month fixed effects to account for seasonality. The vector Zivt contains time-varying household, farm, and village characteristics. The term φ i contains household fixed effects, and ivt is the error term (the standard errors are clustered at the village level). The expression 1it (t = 2012) equals 1 for the second round and is 0 otherwise. The parameter δ 1d is the effect of living less than d km away from a road, once it has been upgraded. The parameter δ 0d provides a test of the common trends assumption over the period 2008–2010. However, work had already started during the 2010 survey and may have generated disturbances. If this is the case, we will wrongly reject the common trends assumption. Our evaluation of the program primarily compares 2008 and 2012. 5.2.2. Matching First, we select households based on their distance to the nearest major road. We thus create “buffers” around roads that define which households might act as controls. For instance, if we assess the effect of living less than 30 km away from an upgraded road, the households in the control group also live less than 30 km from a major road. With this set of potential controls defined, we perform a standard matching based on observable characteristics. The matching is done at the household level, since it allows us to use household characteristics (ob- served in 2008) as predictors. In addition, we use the government selection criteria and include the fol- lowing variables: quality of the closest road prior to 2008, distance to the nearest border crossing, to the nearest major city, to the nearest district headquarters, and to the nearest market. We also use area characteristics, such as population density, agroecological zones, and labor-market activity. Table S2.3 presents summary statistics of these variables. We use an Epanechnikov kernel-matching procedure based on the estimated propensity scores. We choose this non-parametric matching approach because it uses more information to construct the coun- terfactual (Caliendo and Kopeinig 2008). Compared to specification (1), it simply consists in running the household fixed effects regression with the kernel weights obtained from the matching logit. 5.2.3. Price Estimations Prices are collected at the community level, and we focus on rural areas for consistency. The specification is similar to (1) except that observations and matching are done at the village level. We use the same set of predictors, and the household-level regressors are averaged at the village level. 13 Figure S2.6 shows the composition of the treated group, by distance to the road, for each of these thresholds d. 114 Dumas and Játiva 6. Results 6.1. Matching Results Table S2.4 presents the results of the propensity score estimation. The matching is performed for each distance threshold, but table S2.4 only reports the results for the 20 km threshold, as an example. We Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 find that georeferenced controls, as well as road and village characteristics, are good predictors of the treatment. Several results indicate that the program guidelines were followed. First, households living closer to a gravel major road are more likely to have that road upgraded than those living near soil roads, followed by those living near paved and sealed roads. As stated earlier, only roads in poor condition were selected to be improved, and infrastructure projects related to maintenance were given first priority. These results indicate that roads that were in poor condition but did not demand large investments were selected first. Furthermore, the closer a household is to a district headquarters, the higher its probability of having a road upgraded, in line with the government’s objective. A few characteristics that may indicate higher living standards seem to have a negative effect on the probability of treatment, namely, more durable assets, a higher share of households with at least one member employed in wage labor. For each matching, we find a suitable control group. We present as examples the common support assessment for the 20 km and 30 km matchings (figs S2.7 and S2.8). Figure S2.9 displays the spatial locations of treated and control households on the common support, as well as the discarded observations. Table S2.5 shows that some imbalances remain after the matching in the control variables. More wor- risome, panel B displays even starker differences between treated and control groups in the 2008 outcome variables. Households in the treated area produce more rice and less maize. Rice prices are higher in the treated area, and treated households consume less rice than the control households. While we may ques- tion the utility of the matching in this context, we argue that the control group with the matching is more comparable to the treatment group than without the matching. Ultimately, we test the common trends assumption between 2008 and 2010 and find fewer invalidations with the matching than without. We therefore prefer the matching specification and present its results. 6.2. Main Results Migration and attrition We begin by assessing whether roads triggered more migration. This could be a potential threat to our identification strategy if better roads induce selective attrition in the panel. In addition, it is interesting to assess whether better roads induce changes in how households obtain income. However, traveling along a road of poor quality is unlikely to represent a substantial share of migration costs, and we expect that the program has little effect on migration. The next figures display the estimates for each treatment cutoff (ranging from living less than 20 km away from the road to living less than 50 km away from the road). The provided estimates are the household fixed effects with propensity score matching. The confidence interval is at the 90 percent level. Figure 2(a) provides the effect of the treatment on the likelihood that the household is not observed in the last round of the panel. The average of this variable in the sample of control households living less than 50 km away from a road is 0.01, and the number of observations used for the estimations ranges from 1,127 households living less than 20 km away to 1,551 households living less than 50 km away.14 The figure shows that there is little evidence that better roads induce mi- gration/attrition of a complete household. Household attrition would be the most problematic for us, as outcomes are measured at the household level, not at the individual level. We also assess whether the household modifies its structure. Figure 2(b) shows that there is no attrition of adults due to the treat- ment, while fig. 2(c) finds that children are less likely to move out of the household when they are treated. However, when directly assessing whether household composition is affected by the road upgrade, we find no change (fig. S2.10). This suggests that the household is also less likely to receive new members. Last, 14 Table S2.6 displays the number of observations separately by control and treatment groups. The World Bank Economic Review 115 Figure 2. Effect of Road Rehabilitation by Distance to the Road. Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 Source: Authors, based on LSMS-ISA 2008–2009, 2010–2011, and 2012–2013 and TANROADS data. Note: The figures show estimates of a fixed-effects strategy with propensity score matching; FEs are at the household level for all variables except prices, where FEs are at the village level. Confidence intervals are at the 90 percent level. The number of observations in each figure has the following ranges: 2(a): 1,127–1,551; 2(b): 2,890–3,972; 2(c): 2,619–3,742; 2(d): 1,645–3,411; 2(e): 304–503; 2(f): 169–312; 2(g): 260–436; 2(h): 303–504; 2(i): 1,517–3,086; 2(j): 1,517–3,086; 2(k): 1,491– 3,024; 2(l): 4,065–5,676; 2(m): 1,517–3,086; 2(n): 1,517–3,086; 2(o): 1,491–3,024; 2(p): 4,065–5,676; 2(q): 4,065–5,676; 2(r): 4,065–5,676; 2(s): 4,054–5,665; 2(t): 1,645–3,411; 2(u): 1,645–3,411; 2(v): 1,645–3,411; 2(w): 951–2,104; 2(x): 1,096–1,719. . Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 Dumas and Játiva Figure 2 – continued 116 The World Bank Economic Review 117 we assess whether temporary migration becomes possible after the road upgrade. Figure 2(d) depicts the effect of the treatment on the number of household members away for at least one month in the previ- ous year. Figure S2.11(a) provides the test of the common trends assumption (i.e., the δ 0d coefficients). It shows that the common trends assumption might be violated in the instance of the temporary migration variable: households living less than 34 km away from a road display a higher likelihood of migrating, Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 and this likelihood increased as early as 2010. Henceforth, for the sake of brevity, we will only refer to the tests of the common trends assumption when they seem to be rejected, but the full set of results is available in fig. S2.11. We will first explore whether, as in Casaburi, Glennerster, and Suri (2013), changes in the price system occur as a consequence of the upgrade and, if so, whether and how households adjust to this new price system. Prices As only one observation per village is reported for prices and only for available goods, we focus on commonly exchanged goods. Rice and maize are by far the most traded products, and they are the products for which we have sufficient observations to conduct the evaluation. Figure 2(e) displays a clear decline in the price of rice, for all treatment thresholds, and the decrease is very substantial compared to the mean. The point estimates for maize also tend to be negative (fig. 2(f)), but the coefficients are at the margin of significance. A decrease in prices would have a negative impact on welfare for net producers and a positive impact on welfare for net consumers. Insofar as rice and maize are commonly produced by Tanzanian farmers, this decrease in rice prices could reduce a household’s utility. We then check whether prices are generally lower, which would limit the negative effect on welfare for producers. Figures 2(h) and 2(g) show that the prices of sugar and of kerosene are not negatively impacted by road improvement.15 Agricultural outcomes and consumption We now assess whether the price changes are associated with the effects of roads on harvest, sales, land area, and consumption for rice and maize. Figure 2(i) shows that rice harvest and rice area are lower for all treatments (rice sales are also lower but non-significant), but consumption does not change.16 All the variables are expressed in kilograms and not in values. This is consistent with the following mechanism: as the rice price is lower, households reduce their rice production and fail to exploit the decrease in rice prices because they do not consume more rice. As rice is a luxury good, consumption may remain the same when its price decreases as producer income also decreases. We also find changes in the maize sector (increase in production, decrease in consumption) but the common trends tests on maize variables suggest that we should interpret these effects with caution. As mentioned earlier, the maize price is globally increasing over the period, and figs S2.11(j) to S2.11(l) suggest the possibility that forthcoming treated households increase their maize production in the first years of the panel. If anything, this plays “against” the mechanism we suggest of increased competition on prices along the road. Finally, given the price decline and reduction in land area for rice, we should observe an agricultural in- come reduction, unless the households manage to substitute rice with other high-return crops. Figure 2(w) suggests that agricultural income decreases and the decrease is substantial (−40 percent) for some thresh- olds (around 40 km). Using the agricultural income average in the control group (TSh 315,523), this would mean a loss of TSh 126,210 for the most impacted households. 15 Overall prices are higher in treated areas in 2010. 16 Total land area does not change, suggesting that households do not fully withdraw from agriculture but rather switch to other crops. 118 Dumas and Játiva We now turn to other outcome variables to evaluate whether the upgraded road brings improvements in other dimensions. Assets and Welfare First, we assess whether households deplete their assets to cope with the shock. Figures 2(q) and 2(r) plot Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 the impact of rehabilitated roads on durable and productive asset ownership. Both variables are built from a principal component analysis run on a series of durable goods ownership (consumption goods for the first, productive assets for the second). The effect on durable assets is large and negative, but fig. S2.11(n) suggests that the decline could have started earlier than 2012, although the coefficients are not significant. The effect of roads on productive assets is also negative but usually not significant. We also use a general and subjective welfare measure to assess whether households valued their access to a better road. The subjective welfare measure is based on a question with a 7-point response scale asked of household members, and we average their answers.17 The use of a subjective welfare question has limitations but is useful in a setting where prices change, and aggregate consumption measures can- not be made comparable over time (Atkin, Faber, and Gonzalez-Navarro 2018). Figure 2(s) shows that households with an upgraded road do not report lower life satisfaction. Labor market To better understand how households adjust their activity, we turn to the information on the labor market. The main question is whether the production changes are associated with changes in labor supply and demand. Figure 2(v) confirms that treated households reduce their on-farm labor, and fig. 2(t) identifies that they substantially increase (+50 percent) wage work (at least one household member works for a wage). The data do not allow us to identify where individuals work. We do not find an increase in hired labor on farms in the village (fig. 2(u)), so they are likely to be hired in other sectors.18 It would be interesting to identify whether wage workers are hired in the village or outside the village. We also find an increase in wage income (fig. 2(x)), also approximately 40 percent at the maximum. Again, using the average in the control group (TSh 192,815), this would imply an increase in wage income of TSh 77,126, which does not compensate for the loss in agricultural production. Interestingly, the effects on income seem to materialize only when accounting for households who are farther away from the road. One explanation could be that they are impacted by goods competition without being able to exploit the improved road. 6.3. Heterogeneity of the Effect Overall, we find that exposure to a rehabilitated road is associated with a reduction in on-farm activity and an increase in wage work. We now wish to assess whether the effects of the road program are hetero- geneous with respect to household characteristics, in particular with respect to being a rice producer and with respect to the distance to the border. Our expectation would be that rice producers should be more affected than others and that households who are closer to the border are more affected, since cheap rice is imported from neighboring countries. Table S2.7 shows that the program led to a decrease in satisfaction, specifically for rice producers, but an increase in satisfaction for net consumers. We also find that net producers increase wage labor, which results in an increase in wage income. Additionally, households living close to the border experience a large decrease in satisfaction, as well as a reduction in agricultural income. 17 The exact wording of this question is “How satisfied or dissatisfied would you say you are with your life?” 18 A complementary analysis in Dumas and Játiva (2020) on the Labor Force Survey confirmed these results and showed in addition that the sectors in which individuals were finding wage work were the hospitality industry (e.g., restaurants), construction, and transport, which appears consistent with expected effects of a road improvement. The World Bank Economic Review 119 While these results should be taken with caution as they rely on a small number of observations, they are suggestive evidence for the mechanisms we propose. 6.4. Robustness and Placebo Tests We first assess whether our results hold in a more standard difference-in-difference framework, without Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 matching. We define two treatments: households living 20 to 30 km away from an upgraded road and those living 31 to 50 km away. We expect the second group to be less affected than the first. Table S2.8 shows that we find no effect for the second treatment. The main effects when the households live less than 30 km away are confirmed with strong negative impacts on rice prices, sales, harvest, and land area. We also provide tests of other possible effects of roads. We do so in a more compact way by simply providing estimates with treatment being “less than 20 km away,” “less than 30 km away,” and “less than 40 km away” from an upgraded road. We find a few impacts of the program on the implementation of community projects, no changes in security, nor changes in agricultural practices, such as the use of pesticides and inorganic fertilizer (table S2.9). We also checked whether households had access to a greater variety of goods (Gunning, Krishnan, and Mengistu 2018; Aggarwal 2018). We use as an outcome variable the number of goods for which prices are collected in the market and do not find any change. We also do not find that the number of sectors that hire labor has increased. We then run the following placebo tests to assess the validity of our estimation strategy. First, we provide tests of the common trends assumption over longer periods: from 2003 to 2010 using a wealth index built in the DHS (table S2.10), and from 1988 to 2012 using a 10 percent sample drawn from the census (table S2.11). We do not detect differences in wealth trends between treated and control areas. We also use roads that were upgraded after 2013 as placebo roads. We define a placebo treatment for roads with a feasibility study by 2014 and compare them to roads that were neither upgraded in 2012 nor had undergone a feasibility study by 2014. Table S2.12 shows that most coefficients are not significant, suggesting that the allocation of upgrades is not endogenous.19 We now assess the robustness of our results to sample changes. We first restrict control villages to regions that did not receive any upgrade to avoid contamination bias, for instance due to general equilib- rium effects. Table S2.13 shows that the main effects remain, but we often lose significance. Given that the main results survive when we account for potential general equilibrium effects at the regional level, we are inclined to conclude that such large effects are not present. However, we acknowledge that, as usual, a difference-in-difference strategy is not well positioned to identify general equilibrium effects. Second, we also exclude the villages that are less than 5 km from cities with more than 5,000 inhabitants because there might be endogenous selection of the treated cities (Banerjee, Duflo, and Qian 2020). Table S2.14 confirms the effects on rice cultivation and asset depletion. We then assess two alternative treatment definitions. First, we interpret the negative effect of the road as due to increased competition. However, it is unclear in which market such competition occurs. Indeed, our treatment could proxy for another treatment such as being close to a city (i.e., a market) that benefits from the road upgrade. We define a new treatment, which is being located less than d km away from a city, which is treated in the sense that the program improved its connectivity to the rest of the country or to neighboring countries. Figure S2.12 displays the cities that have been identified as benefitting from the program. Globally, the effects operate in the same direction but are lower and often lose significance (table S2.15). Our previous treatment variable therefore appears to have stronger predictive power than the distance to a treated city. We thus conclude that proximity to a road triggers the observed changes rather than proximity to a treated city. This suggests that the main markets in which the households face increased competition are local rather than located in major cities. 19 The coefficients that are significantly different from 0 (e.g., rice sales) are actually of the opposite sign to those obtained in our main results and therefore cannot explain our findings. 120 Dumas and Játiva Last, throughout our analysis we have used the as-the-crow-flies distance to the road. If individuals have to use specific secondary roads to reach a major road, this distance measure may not reflect the true transportation cost between the village and the road. We use the network of secondary roads from the open-source routing service OpenStreetMap (OSM);20 fig. S2.13 displays the secondary roads in these data and computes the travel time between two points.21 Table S2.16 shows that most of the effects Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 remain with this alternative specification, except for income effects. 7. Conclusion We evaluate the effects of road infrastructure upgrades on the labor and product market participation, prices, and income of rural households in Tanzania. We combine household fixed effects with propensity score matching. Our results suggest damaging effects of road improvements on the rural population: we observe a decrease in the rice price and a reallocation of labor away from farm work. The increase in wage work does not seem to compensate for the loss in agricultural income (but the identification of this effect is not robust across specifications). This fits with predictions obtained from trade models where farming households face competition from lower priced goods and with previous results on limited sectoral reallocation. While our paper is one of the few that indicate negative effects of transportation infrastructure, the results should not be extrapolated to other transportation investments for the following reasons. First, our results are confined to the effect of major roads, not secondary roads. Beuran, Gachassin, and Raballand (2015) provide several arguments why improved access to a major road may not deliver the expected positive effects. They stress that accessing a main road may remain difficult and that transportation ser- vices may be lacking; therefore, households may be unable to exploit the improved access to cities. In our data set, the majority of the households live less than 5 km away from a rural road, which then connects to a major road. This is both near and far. Five kilometers would not prevent bags of goods from being delivered but would certainly limit labor-market opportunities offered by the improved major road since it remains quite distant. Second, the results may be specific to this program. Indeed, it fulfilled several objectives: not only poverty reduction but also improvements in Tanzania’s connectivity with neighboring countries. In our study, we are unable to identify the exact source of competition because we do not have information on trade flows. The results are however consistent with competition associated with cheap rice tran- siting through neighboring countries. We acknowledge that, had the choice of roads been different or had government trade policy been different, the welfare effect on rural households could also have been different. Last, the intervention we evaluate is a road rehabilitation program, not a road construction program. As previously discussed, we expect a stronger impact in the short run. Our estimates are also only short-run effects. It is therefore indicative of how improved connectivity, when trade networks are already present, may increase competitive pressure on rural households. Our focus on rehabilitated rather than newly built roads may explain why we find negative effects. For newly built roads, the effects may take time to materialize since trade networks are reorganized, but this also leaves time for households to adjust. We cannot exclude the possibility that, in the longer run, producers will adjust to the new economic environment. However, it could also be the case that the shock affects them durably. A longer run study would be needed to make a judgment on this question. 20 OSM maps were obtained from http://download.geofabrik.de in February 2020. OSM road data for 2008 cannot be retrieved. 21 We use the Stata command osrmtime (Huber and Rust 2016). The speed profiles are clearly optimistic: the default speed used for a car on a secondary road is 55 km/h, while Aggarwal et al. (2018) estimate that speed on secondary roads averages 22 km/h. The World Bank Economic Review 121 Several questions remain: First, we are limited in our assessment of the program by the fact that the maize price is not stable over the period. Given that maize is one of the major products of Tanzania, it would be good to reproduce the research in a more stable economic environment. Second, it seems intuitive to hypothesize that the road has a deeper effect on the goods market (in the sense that it has a more profound outreach) than on the labor market because individuals would need to access the road on Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 a regular basis. Potentially, this could be estimated via heterogenous impacts by distance to the road. We have attempted to do so, but a larger sample would be required to identify the distance effects. This is also crucial in the design of the optimal investment in roads to understand these effects. Third, even if we identify that individuals opt more often for wage work, we have few details on these wage activities. It is therefore difficult to identify whether they existed before the road upgrade (but households did not opt for them) or if they opened with the road upgrade. In the longer run, the effects would be vastly different if individuals were taking lower productivity jobs out of distress or if these wage activities were rather a progressive structural transformation of rural areas, for which productivity may increase. Data Availability Statement The roads data are available upon request while the LSMS-ISA data are publicly available. References Aggarwal, S. 2018. “Do Rural Roads Create Pathways Out of Poverty? Evidence from India.” Journal of Development Economics 133(C): 375–95. Aggarwal, S., B. Giera, D. Jeong, J. Robinson, and A. Spearot. 2018. “Market Access, Trade Costs, and Technol- ogy Adoption: Evidence from Northern Tanzania.” NBER Working Paper 25253, National Bureau of Economic Research, November: 1-59. Asher, S., and P. Novosad. 2020. “Rural Roads and Local Economic Development.” American Economic Review 110(3): 797–823. Atkin, D., B. Faber, and M. Gonzalez-Navarro. 2018. “Retail Globalization and Household Welfare: Evidence from Mexico.” Journal of Political Economy 126(1): 1–73. Attanasio, O., P. K. Goldberg, and N. Pavcnik. 2004. “Trade Reforms and Wage Inequality in Colombia.” Journal of Development Economics 74(2): 331–66. Baffes, J., V. Kshirsagar, and D. Mitchell. 2017, 05. “What Drives Local Food Prices? Evidence from the Tanzanian Maize Market.” World Bank Economic Review 33(1): 160–84. Banerjee, A., E. Duflo, and N. Qian. 2020. “On the Road: Access to Transportation Infrastructure and Economic Growth in China.” Journal of Development Economics 145: 102442. Beuran, M., M. Gachassin, and G. Raballand. 2015. “Are There Myths on Road Impact and Transport in Sub-Saharan Africa?” Development Policy Review 33(5): 673–700. Bird, J., and S. Straub. 2020. “The Brasília Experiment: The Heterogeneous Impact of Road Access on Spatial Devel- opment in Brazil.” World Development 127: 104739. Burgess, R., R. Jedwab, E. Miguel, A. Morjaria, and G. Padró i Miquel. 2015. “The Value of Democracy: Evidence from Road Building in Kenya.” American Economic Review 105(6): 1817–51. Caliendo, M., and S. Kopeinig. 2008. “Some Practical Guidance for the Implementation of Propensity Score Match- ing.” Journal of Economic Surveys 22(1): 31–72. Casaburi, L., R. Glennerster, and T. Suri. 2013. “Rural Roads and Intermediated Trade: Regression Discontinuity Evidence from Sierra Leone.” Castañeda, A., D. Doan, D. Newhouse, M. C. Nguyen, H. Uematsu, and J. P. Azevedo. 2018. “A New Profile of the Global Poor.” World Development 101: 250–67. Chandra, A., and E. Thompson. 2000. “Does Public Infrastructure Affect Economic Activity?: Evidence from the Rural Interstate Highway System.” Regional Science and Urban Economics 30(4): 457–90. Chaurey, R., and D. T. Le. 2021. “Rural Infrastructure Development and Economic Activity.” 122 Dumas and Játiva Currie, J., and A. E. Harrison. 1997. “Sharing the Costs: The Impact of Trade Reform on Capital and Labor in Morocco.” Journal of Labor Economics 15(3): 44–71. de Janvry, A., M. Fafchamps, and E. Sadoulet. 1991. “Peasant Household Behaviour with Missing Markets: Some Paradoxes Explained.” Economic Journal 101(409): 1400–17. Dercon, S. 1998. “Wealth, Risk and Activity Choice: Cattle in Western Tanzania.” Journal of Development Economics Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 55(1): 1–42. Dercon, S., D. O. Gilligan, J. Hoddinott, and T. Woldehanna. 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. Donaldson, D. 2018. “Railroads of the Raj: Estimating the Impact of Transportation Infrastructure.” American Eco- nomic Review 108(4–5): 899–934. Dumas, C., and X. Játiva. 2020. “Better Roads, Better Off? Evidence on Upgrading Roads in Tanzania.”. Working Paper 518, Faculté des SES, Université de Fribourg. FAO. 2018. “Small Family Farms Country Factsheet.” Food and Agriculture Organization of the United Nations (FAO) Country Factsheets: 1–2. ———. 2020. “FAOSTAT Database.” Fiorini, M., and M. Sanfilippo. 2022. “Roads and Jobs in Ethiopia.” World Bank Economic Review 36(4): 999–1020. Fiorini, M., M. Sanfilippo, and A. Sundaram. 2021. “Trade Liberalization, Roads and Firm Productivity.” Journal of Development Economics 153: 102712. Ghani, E., A. G. Goswami, and W. R. Kerr. 2016. “Highway to Success: The Impact of the Golden Quadrilateral Project for the Location and Performance of Indian Manufacturing.” Economic Journal 126(591): 317–57. Government of Tanzania. 2008. “10 Year Transport Sector Investment Programme (TSIP).” 1–119. ———. 2011. “The Tanzania Five Year Development Plan 2011/2012-2015/2016.” Technical report, President’s Of- fice, Planning Commission. Gunning, J. W., P. Krishnan, and A. T. Mengistu. 2018. “Fading Choice: Transport Costs and Variety in Consumer Goods.” CEPR Discussion Papers 12910. Harrison, A., and G. Hanson. 1999. “Who Gains from Trade Reform? Some Remaining Puzzles.” Journal of Devel- opment Economics 59(1): 125–54. Heckman, J., and S. Navarro-Lozano. 2004. “Using Matching, Instrumental Variables, and Control Functions to Estimate Economic Choice Models.” Review of Economics and Statistics 86(1): 30–57. Huber, S., and C. Rust. 2016. “Calculate Travel Time and Distance with OpenStreetMap Data Using the Open Source Routing Machine (OSRM).” Stata Journal 16(2): 1–8. Jedwab, R., and A. Storeygard. 2021. “The Average and Heterogeneous Effects of Transportation Investments: Evi- dence from Sub-Saharan Africa 1960–2010.” Journal of the European Economic Association 20(1): 1–38. Khandker, S. R., Z. Bakht, and G. B. Koolwal. 2009. “The Poverty Impact of Rural Roads: Evidence from Bangladesh.” Economic Development and Cultural Change 57(4): 685–722. McCullough, E., C. Zhen, S. Shin, M. Lu, and J. Arsenault. 2022. “The Role of Food Preferences in Determining Diet Quality for Tanzanian Consumers.” Journal of Development Economics 155: 102789. Minot, N. 2010. “Staple Food Prices in Tanzania.”. Minten, B., B. Koru, and D. Stifel. 2013. “The Last Mile(s) in Modern Input Distribution: Pricing, Profitability, and Adoption.” Agricultural Economics (United Kingdom) 44(6): 629–46. Nakamura, S., T. Bundervoet, and M. Nuru. 2020. “Rural Roads, Poverty, and Resilience: Evidence from Ethiopia.” Journal of Development Studies 56(10): 1838–55. Qin, Y., and X. Zhang. 2016. “The Road to Specialization in Agricultural Production: Evidence from Rural China.” World Development 77(C): 1–16. Revenga, A. 1997. “Employment and Wage Effects of Trade Liberalization: The Case of Mexican Manufacturing.” Journal of Labor Economics 15(3): 20–43. Rudolf, R. 2019. “The Impact of Maize Price Shocks on Household Food Security: Panel Evidence from Tanzania.” Food Policy 85(May): 40–54. Schwab, K., and M. E. Porter., eds. 2008. The Global Competitiveness Report 2008–2009. Geneva: World Economic Forum. The World Bank Economic Review 123 Schwab, K., and X. Sala-i Martin., eds. 2016. The Global Competitiveness Report 2016–2017. Geneva: World Eco- nomic Forum. SERA. 2012. “Time to Re-think the Food Crops Export Ban.” Policy Brief 1, USAID. ———. 2016. “Rules-Based Transparent System for Emergency Food Imports.” Policy Brief 7, USAID. Shamdasani, Y. 2021. “Rural Road Infrastructure & Agricultural Production: Evidence from India.” Journal of De- Downloaded from https://academic.oup.com/wber/article/39/1/104/7647241 by WORLDBANK THIRDPARTY user on 05 February 2025 velopment Economics 152: 102686 A Shrestha, S. 2020. “Roads, Participation in Markets, and Benefits to Agricultural Households: Evidence from the Topography-Based Highway Network in Nepal.” Economic Development and Cultural Change 68(3): 839–64. Singh, I., L. Squire, and J. Strauss. 1986. Agricultural Household Models: Extension, Application and Policy. Johns Hopkins University Press, Baltimore, MD. Smith, J. A., and P. E. Todd. 2005. “Does Matching Overcome LaLonde’s Critique of Nonexperimental Estimators?” Journal of Econometrics 125(1–2): 305–53. 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. World Bank. 2008. “Safe, Clean, and Affordable...Transport for Development: The World Bank Group’s Transport Business Strategy for 2008–2012.” Technical report. Topalova, P. 2010. “Factor Immobility and Regional Impacts of Trade Liberalization: Evidence on Poverty from India.” American Economic Journal: Applied Economics 2(4): 1–41. Volpe, Martincus, C., J. Carballo, and A. Cusolito. 2017. “Roads, Exports and Employment: Evidence from a Devel- oping Country.” Journal of Development Economics 125(C): 21–39.