WPS8548 Policy Research Working Paper 8548 Heterogeneous Impacts of Main and Feeder Road Improvements Evidence from Ethiopia Atsushi Iimi Haileyesus Mengesha James Markland Yetmgeta Asrat Kefargachew Kassahun Transport and Digital Development Practice August 2018 Policy Research Working Paper 8548 Abstract Rural access is among the most important infrastructure output market access was improved by feeder road improve- constraints in rural Africa. Using the results from compre- ment. In addition, the household’s nonagricultural income hensive household surveys and other data from Ethiopia, is somehow increased by improved road connectivity. There the paper recasts light on the heterogeneous impacts of road must be secondary effects. The transport demand function accessibility on agriculture and nonagricultural growth. It is estimated with additional data indicates that as the road found that crop production is increased by major and feeder network improves, people’s mobility increases. Furthermore, road improvements. Significant synergy is also found. When local business employment is found to increase with road investigating further into this effect, there are two impacts: improvements. To meet the increasing demand for mobility, farmers’ access to the input market, especially fertilizer, was efficiency and frequency of transport services are important. improved mainly by major corridor improvement. And This paper is a product of the Transport and Digital Development Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted at aiimi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Heterogeneous Impacts of Main and Feeder Road Improvements: Evidence from Ethiopia Atsushi Iimi ¶ † Haileyesus Mengesha † James Markland † Yetmgeta Asrat § Kefargachew Kassahun § ‡ Key words: Rural access, Agricultural production, Transport service demand, Local job creation. JEL classification: R41; Q12; C33; C34. ¶ Corresponding author † Transport & Digital Development Global Practice, World Bank Group § Ethiopian Roads Authority ‡ We would like to acknowledge both TPO consultant and URRAP Impact Assessment teams for their efforts to carry out various surveys and data collection. -2- I. INTRODUCTION Rural access is among the most important infrastructure constraints to development in rural Africa. A spatial analysis by Dorosh et al. (2012) indicates that crop productivity in Africa is closely related to transport connectivity. Dercon et al. (2012) show that access to an all- weather road has a significant impact on farmers’ welfare in Ethiopia. Similar results were found by Bachewe et al. (2017). In general, although causality is always a difficult issue, the literature is supportive of a wide range of socioeconomic impacts of improved rural accessibility around the world. In the short term, transport costs and travel time can be reduced by improved road conditions, and resilience to drought is improved (e.g., Lokshin and Yemtsov 2005; Danida 2010). Over the longer term, agricultural productivity would likely be increased (Khandker et al. 2009; Bell and Van Dillen, 2014), and firms would become more profitable with more jobs created (Mu and van de Walle, 2011). Poverty could then be alleviated (Dercon et al. 2009; Khandker and Koolwal, 2011). However, there are still divergent views on results chains for rural access improvement. For example, there has long been a debate on the relative importance of agricultural and non- agricultural impacts. Bell and van Dillen (2014) show that a higher level of road accessibility increased crop output prices by 5 percent in India. Improved rural road connectivity can also increase agricultural employment and working hours in Nicaragua. (Rand, 2011). According to Diao et al. (2012), agricultural growth has a greater impact on poverty reduction in many African countries. This is primarily because the vast majority of the poor engage in agricultural production in rural Africa. However, there are also spillover effects on non-agricultural sectors. Manufacturing growth, especially in the agro-processing subsector, is often found to be critical for poverty reduction (Dorosh and Thurlow, 2017). Some studies, such as Shirley and Winston (2004) and Datta (2012) show that improved road accessibility can reduce firms’ inventory costs, though the studies are focused on connectivity to primary roads. But firms are generally likely to flock together where good transport infrastructure is available (e.g., Cieślik and Ryan, 2004; -3- Boudier-Bensebaa, 2005). New local businesses may also emerge (Mu and van de Walle, 2011). Given the rapid urbanization trend, non-agricultural activities, regardless of whether they are formal or informal, are certainly becoming more and more important even in Africa. The current paper examines the question of whether improved rural accessibility has promoted agricultural or non-agricultural growth, or both. It attempts to identify the relative importance of different growth channels. For the agricultural growth channel, it investigates whether the road improvements impacted on people’s access to markets for inputs or outputs. The literature is supportive of both cases. While proximity allows farmers to use more advanced inputs, such as fertilizer (e.g., Sheahan and Barrett, 2017; Damania et al., 2017), farmers’ participation in the output market would be stimulated by lower crop transport costs (e.g., Key et al., 2000). For the non-agricultural growth channel, the paper investigates further the causes for non-agricultural income growth. The analysis will be extended to the estimation of people’s demand for transportation services and the indirect impacts on local businesses, especially, job creation. Of particular note, the paper examines the impacts of different types of road improvement works. Though often forgotten, different roads fulfill different functions. While feeder roads provide last-mile connectivity to local farmers and businesses, the higher level of the road network, such as trunk roads and regional corridors, connects major cities and economic activities, carrying consolidated traffic between urban and rural areas. There is little evidence of the interaction effects between them. One of the few studies is Bell and van Dillen (2014), which indicate some heterogeneous impacts between highways and feeder roads: Crop prices decrease with distance along a major road link but increase with distance along a stretch of feeder road. By contrast, Khandker et al. (2009) show that rural road investment can increase output prices, while reducing farmgate prices of inputs. In the general equilibrium context, Warr (2008) shows the particular importance of the minimum level of accessibility to reduce poverty in the Lao People’s Democratic Republic. By contrast, Escobal and Ponce (2002) find that households in villages receiving motorable roads had an average increase in -4- household income of 35 percent, with no effect for treated households served by non- motorable roads. The paper will cast light on such heterogeneity and complementarities across different types of roads. Our data come from five rounds of household surveys carried out in Ethiopia for the period 2012 to 2016. In recent years, the Government of Ethiopia has been investing heavily in both primary and feeder roads. Under the Road Sector Development Program (RSDP), many Trunk and Link Roads have been rehabilitated. There are also rural road initiatives, such as Universal Rural Road Access Programme (URRAP), which is focused on feeder road improvement at the village (kebele) level. Different levels of connectivity are likely to contribute to the economy in different ways. The paper is organized as follows: Section II provides a brief overview of the country context, specifically focused on the agriculture and road sectors. Section III develops our empirical strategy. Section IV provides more details on our data. Section V presents main results and discusses policy implications. Then Section VI concludes. II. AGRICULTURE AND ROAD SECTORS IN ETHIOPIA In Ethiopia, agriculture remains among the most important economic sectors. It produces about one-third of GDP, employs 70 percent of the workforce and accounts for 80 percent of the country’s merchandise exports (Ethiopian Agricultural Transformation Agency, 2014). Ethiopia produces about US$6 billion of crops a year, including teff, maize, coffee, beans, sorghum and wheat. Coffee is one of the traditional export crops of the country. Agricultural productivity is generally low but compares favorably to neighboring countries. Earlier studies (e.g., Taffesse et al., 2012; Spielman et al., 2012) show that agricultural growth was largely attributed to expansion of land area cultivated during the 1990s, but productivity growth has also been contributing since the early 2000s. -5- Still, the use of advanced inputs, such as fertilizer and improved seeds, remains limited. Fertilizer was applied to less than 40 percent of cereal acreage, and pesticides to about 20 percent of land. The use of improved seeds and irrigation is limited to several percent on average (Taffesse et al. 2012). The poor rural road network has long been recognized as a crucial constraint preventing the timely delivery of agricultural inputs. Transport costs are estimated to account for 64-80 percent of fertilizer farmgate prices (Rashid et al. 2013). In Ethiopia, there are more than 110,000 km of roads, of which only 14,000 km or about 13 percent are paved. In recent years, the government has been making significant efforts to develop both main corridors and rural feeder roads. About 26,000 km of federal roads are generally well-maintained. In 2010, the Universal Rural Road Access Program (URRAP) was embarked upon, aimed at connecting all communities (kebeles) by all-weather roads. Under the program, about 46,000 km of rural roads have been rehabilitated, and more than 5,600 villages (kebeles) were connected. The total costs in the first 5 years of the program amounted to 28 billion Ethiopian birr (ETB) or about US$1.4 billion. Still, there are a number of feeder roads that are in poor condition. Only about 29 percent of the total road network is good condition in Ethiopia.1 The paper is focused on assessing the impacts of rehabilitation works along four major roads that were rehabilitated under the Road Sector Development Program (RSDP) and a number of feeder roads that were simultaneously improved under a separate program, URRAP, in the same areas. The four main roads are: (i) Aposto-Wendo-Negele Road (268 km), (ii) Mekenajo-Dembidolo Road (181 km), (iii) Kombolcha-Bati-Mille Road (133 km), and (iv) Ankober-Awash Arba Road (89 km). While the first two roads are located in Oromia and Southern Nations, Nationalities and Peoples’ Regions, the last two connect Amhara and Afar Regions (Figure 1). 1 See Nakamura (2017) for more details. -6- Figure 1. Road network in Ethiopia Source: World Bank (2016) based on Ethiopia Road Authority data. III. EMPIRICAL STRATEGY To assess the impacts of road improvements, the following production function is considered: ln Yit j   0   1 DitRSDP   2 Dit URRAP   3 DitRSDP  Dit URRAP   k  k ln X ikt   m  m Z imt  ci  ct  u it (1) where Yitj represents the amount of household i’s income from source j at time t. Three different income sources are considered: j = {crop, livestock, non-agriculture}. Agricultural incomes include both own consumption and market sales. On the other hand, remittance is excluded from nonagricultural income. -7- Household income is assumed to depend on two types of road accessibility, DRSDP and DURRAP, as well as agricultural inputs used (X) and other household-specific characteristics RSDP (Z). Dit is set to be one if the main road closest to household i was rehabilitated by the URRAP point of time t. Similarly, Dit is set to be one if the closest feeder road was rehabilitated by URRAP. β3 is expected to capture the interaction effect between the two interventions. As our data are panel, ci and ct are the individual- and time-specific fixed-effects, respectively. Our identification strategy is based on the phased nature of the two road programs. Even along the same road, some segments may have been completed earlier than others. Therefore, some groups benefit from the programs earlier than others. The current paper takes advantage of such time lags in project implementation and applies the difference-in-differences (DID) method with various covariates taken into account. Among the RSDP roads, for example, the Aposto-Wendo-Negele Corridor was divided into three contract lots (Table 1). While two lots were completed by 2013, the final lot was handed over in 2014. The household surveys were carried out in the almost same season every year. This creates an opportunity to distinguish beneficiaries and nonbeneficiaries at a certain point of time. The Kombolcha-Bati-Mille and Ankober-Awash Arba Corridors can be considered as a full comparison group because no road work had yet been completed when the last round of the survey was carried out. Table 1. Timeline of project implementation of the main road works and household surveys Road Lot 2009 2010 2011 2012 2013 2014 2015 2016 2017 Aposto-Wendo-Negele Aposto-Irbamoda Irbamoda-Wadera Wadera-Negele Mekenajo-Dembidolo Mekenajo-Ayira Aiyra-Chanka Chanka-Dembidolo Kombolcha-Bati-Mille Kombolcha-Burka Burka-Mille Ankober-Awash Arba Ankober to Dulesa Dulesa to Awash Arba -8- Feeder road works under the URRAP were also implemented using a similar phased approach. But unlike the above RSDP roads, which are major intercity roads and take 3 to 4 years to be rehabilitated, the URRAP roads are gravel roads. It usually takes 3 to 6 months to carry out improvement works. The completion year differs from road to road even in the same locality (Figure 2). Again, this allows to identify potential beneficiaries and nonbeneficiaries from improved feeder roads. Therefore, depending on the timing, some households may benefit from the RSDP, but not the URRAP, at a particular point of time. An advantage of the DID estimator is that it allows the control of unobserved heterogeneity between the treatment and comparison groups and mitigates the self-selection bias, as far as time-invariant unobservables are concerned. As often discussed, however, this may be a strong assumption. To reduce this risk, time-variant household characteristics are included on the right-hand side of the regression equation (e.g., Jalan and Ravallion 1998). Then, the fixed-effects panel regression is performed. Figure 2. RSDP roads and URRAP roads in project areas (Aposto-Wendo-Negele) (Mekenajo-Dembidolo) -9- (Kombolicha-Bati-Mille) (Ankober-Awasharba) Another important empirical question related to the identification strategy is who the actual beneficiaries are. Unlike other infrastructure services with excludability, roads are generally nonexcludable public goods (except for toll roads). It is not unambiguous whether people who live close to a road are actually benefiting from that road. It must depend on their needs for mobility and destinations of travel. In our data, the average distance that farmers travel to sell their produce at market is 8.9 km, but most farmers travel less than 2 km (Figure 3). Benefits from roads can be both direct, for people who make increased use of a road to travel, or indirect, when school services improve, commodity prices fall or market prices for outputs increase. This inspires an alternative identification strategy based on proximity of households to roads. Although its validity needs to be ex post examined empirically, a threshold of 2km was selected, which is consistent with the people’s travel pattern in Ethiopia as shown above. It also follows the global norms, such as the Rural Access Index (RAI), which measures a proportion of the rural population who live within 2km of an all-season road. To this end, DRSDP and DURRAP in Equation (1) are replaced with DRSDP2km and DURRAP2km, respectively. RSDP 2 km Dit takes the value of one if household i lives within 2km of the closest RSDP road - 10 - URRAP2 km and if that road was rehabilitated at the point of time t. Dit is defined in the same way. To allow this identification strategy, our surveys covered both households who live close to and far from each project road. This identification strategy has the advantage of controlling for unobserved local conditions since we have both beneficiaries and nonbeneficiaries within the same local area along a given road. Though, it is a naïve approach because household placement may be self- selected. It is plausible that those who live closer to the road network may be systematically more productive than remote households. The individual fixed-effects are expected to remove such time-invariant factors. Figure 3. Distribution of distance traveled from households to market .3 .2 Density .1 0 0 10 20 30 Distance to market (km) Under the same identification strategies, two additional models are examined. First, the farmers’ technology adoption decision is analyzed to understand the reasons for possible agricultural income growth. As in earlier studies (e.g., Taffesse et al., 2012; Rashid et al., 2013), access to advanced inputs is a critical constraint in Ethiopia. To this end, the following specification is considered: * ln X ikt   0  1 DitRSDP   2 Dit URRAP   3 DitRSDP  Dit URRAP   k 1  k ln X ikt   m  m Z imt  ci  ct  u it (2) - 11 - Two inputs, k, are considered for X: {Fertilizer (F) and Pesticide (P)}.2 When estimating the equation, one empirical issue is that many input variables are likely to be zero in developing countries. According to Taffesse et al. (2012), fertilizer was applied to less than 40 percent of cereal acreage, and pesticides to about 20 percent of land in Ethiopia. Our sample data are consistent and show even more limited use: About 20 percent and 10 percent of households surveyed used fertilizer and pesticides, respectively. To deal with this problem, the truncated regression model with fixed effects is applied: ln X * jkt  0 if X * ln X jkt   jkt (3) ln  otherwise where ε is a very small positive number to avoid taking the logarithm of zero. Except for the truncated observations, the distribution of the amount of fertilizer or pesticides used is close enough to a normal distribution, which is needed to perform the truncated regression model (Figure 4). Other covariates are the same as the above except for one of the input variables that is now used as a dependent variable. Figure 4. Distributions of fertilizer and pesticides used .4 .4 .3 .3 Density Density .2 .2 .1 .1 0 0 -5 0 5 -5 0 5 10 ln(Fertilizer used, kg) ln(Pesticides used, litre) 2 No data that show the use of improved seeds are available, which is minimal anyway in the current Ethiopian context. - 12 - The other model is the crop sales equation. The road improvements are expected to enhance farmers’ participation in output markets. To this end, the following specification is considered with the crop production value replaced with the total amount of crops sold at market, SALE: * ln SALE it   0   1 DitRSDP   2 Dit URRAP   3 DitRSDP  Dit URRAP   k  k ln X ikt   m  m Z imt  ci  ct  u it (4) Our data cover 28 crops, such as teff, maize, coffee and beans. To aggregate them, producer price data are used from FAOSTAT. Note that as in the case of advanced inputs used, not many farmers have yet participated in market transactions in rural Ethiopia. In our survey data, about one-fourth of households went to market and sold some crops that they produced. Therefore, again, the truncated regression model is used: ln SALE it * if SALE it  0 ln SALE it   (5) ln  otherwise The total value of crop sales at market in logarithm is distributed closely to a normal distribution, which validates the truncated regression model (Figure 5). Figure 5. Distributions of total crop sales .3 .2 Density .1 0 -5 0 5 10 15 ln(Total crop sales, US$) - 13 - IV. DATA The data were collected through five rounds of household surveys carried out by the Ethiopia Road Authority (ERA), along the four main corridors that have been recently rehabilitated or are currently being rehabilitated. As discussed above, they are major Link Roads that share the similar functionality connecting major regional cities to Addis Ababa, the country’s capital. According to the preliminary screening, these four roads share certain common characteristics and are broadly comparable with each other (Table 2). The baseline survey was conducted in 2012 before the first major construction work was started along Aposto-Wendo- Negele Road. Each round of the survey was targeted at the same, approximately 250 households along each main RSDP road. For each road, 15 to 20 villages were selected around the starting, middle and ending point of each road. The villages were also stratified based on distance to the project roads.3 Attrition turned out fairly small over time (Table 3). Table 2. Household characteristics (sample means) by RSDP road Variable Aposto-Wendo- Mekenajo- Kombolicha- Ankober- Negele Dembidolo Bati-Mille Awasharba Size of household 6.00 5.08 5.02 4.24 Dummy variable for male household head 0.81 0.84 0.68 0.74 Age of household head 44.0 44.6 46.3 44.0 Dummy for the level of education attained by household head: 1 to 6 grade 0.27 0.14 0.22 0.13 7 to 8 grade 0.08 0.19 0.06 0.08 9 to 12 grade 0.18 0.19 0.10 0.11 Technical/ vocational certificate 0.02 0.00 0.00 0.01 University/ college diploma 0.06 0.08 0.02 0.03 University/ college degree 0.02 0.04 0.02 0.02 Adult education program/ Non-formal 0.05 0.02 0.05 0.06 Dummy variable for infrastructure services: Electricity for cooking 0.02 0.01 0.02 0.01 Electricity for lighting 0.54 0.78 0.49 0.63 3 The normal cutoff point was originally set to 5km, which is considered to be a normal distance allowing people to walk to a road in Ethiopia. In the actual sample, as shown in the above figure, half of the sample households live less than 2km of the project roads, and another half live 5 to 20km from the roads. - 14 - Table 3. Number of households surveyed Road contract lot Survey year: 2012 2013 2014 2015 2016 Total A1 Aposto-Irbamoda 70 64 68 67 66 335 A2 Irbamoda-Wadera 61 58 58 57 56 290 A3 Wadera-Negele 109 106 106 103 103 527 B1 Mekenajo-Ayira 78 77 78 77 77 387 B2 Aiyra-Chanka 73 73 73 73 73 365 B3 Chanka-Dembidolo 89 86 87 88 87 437 C1 Kombolcha-Burka 138 139 139 136 130 682 C2 Burka-Mille 102 100 99 97 103 501 D1 Ankober-Awash Arba 240 233 227 203 210 1,113 Total 960 936 935 901 905 4,637 The summary statistics indicate that the surveyed households are significantly heterogeneous (Table 4). The household income from crop production ranges from ETB50 to ETB800,000, with an average of ETB9,700 or about US$450. Recall that per capita GDP in Ethiopia was US$706 in 2016. Livestock seems to be as important as crop production in the survey areas. The average income from livestock is about ETB8,000. Many households earn income from other activities, such as forest product sales and seasonal employment. This reflects our sampling strategy: Half of the survey villages were located close to the RSDP roads. Note that remittance income is excluded from our non-agricultural income variable because it is perhaps less relevant to our research context. In the sample, about half of the farmers participate in transactions to sell their produce at market: The average value of crop sales among those farmers is US$573, with a wide variation from only US$1 to over US$130,000. This implies that, not surprisingly, some larger-volume producers are more likely to participate in market sales and earn more than smallholder farmers. For the crop income equation, five inputs are considered for Xk: k = {L, R, I, F,P}, following the agricultural economics literature (e.g., Gyimah-Brempong, 1987; Bravo-Ortega and Lederman,2004). L denotes the number of labor force in each household, which is assumed to include household members between the ages of 16 to 60. While R denotes rain-fed areas used for crop production, I represents irrigated land areas. Two advanced inputs are included: fertilizer (F) and pesticides (P). - 15 - In the livestock income equation, land areas for grazing (G) is used instead of rain-fed or irrigated areas. Fertilizer and pesticides are, of course, irrelevant and removed. To estimate the impacts on nonagricultural income, all these agricultural input variables are excluded. Table 4. Summary statistics Variable Abb. Obs. Mean Std. Dev. Min Max Household income from (Birr): Crop production (including own consumption) YCrop 1,877 9,722 22,162 50.00 800,000 Livestock (including own consumption) YLivestock 1,094 8,072 8,962 40.00 111,000 Nonagricultural activities (exc. remittance) YNonAG 3,157 19,519 30,251 36.00 754,000 Crop sales at market (US$) SALE 1,044 573 4,318 1.03 133,919 Dummy variable for program beneficiaries: RSDP DRSDP 4,639 0.20 0.40 0 1 URRAP URRAP D 4,639 0.67 0.47 0 1 RSDP2mk RSDP within 2km distance D 4,639 0.11 0.31 0 1 URRAP within 2km distance DURRAP2km 4,639 0.33 0.47 0 1 Household labor force (16< age <60) L 4,639 2.64 1.37 0 10 Rain fed crop production area (ha) R 1,721 1.32 2.38 0.025 75.00 Irrigated land area (ha) I 137 0.62 1.14 0.013 10.00 Land area for grazing (ha) G 365 0.57 2.26 0.013 25.00 Fertilizer used (kg) F 927 4.40 17.99 0.005 200.00 Pesticides used (litre) P 441 5.89 57.78 0.003 1200.00 Size of household SIZE 4,639 5.00 2.28 1 14 Dummy variable for male household head MALE 4,639 0.78 0.42 0 1 Age of household head AGE 4,639 45.72 15.00 14 99 Dummy for the level of education attained by household head: 1 to 6 grade 4,639 0.19 0.39 0 1 7 to 8 grade EDU2 4,639 0.11 0.31 0 1 9 to 12 grade EDU3 4,639 0.14 0.35 0 1 Technical/ vocational certificate EDU4 4,639 0.01 0.10 0 1 University/ college diploma EDU5 4,639 0.05 0.22 0 1 University/ college degree EDU6 4,639 0.03 0.17 0 1 Adult education program/ Non-formal EDU7 4,639 0.04 0.19 0 1 Dummy variable for infrastructure services: Electricity for cooking COOK 4,639 0.03 0.18 0 1 Electricity for lighting LIGH 4,639 0.61 0.49 0 1 Time fixed effects: t=2013 T2013 4,639 0.20 0.40 0 1 t=2014 T2014 4,639 0.20 0.40 0 1 t=2015 T2015 4,639 0.19 0.40 0 1 t=2016 T2016 4,639 0.20 0.40 0 1 - 16 - Distance to the nearest RSDP road (km) 4,637 5.79 10.86 0.0003 89.45 Distance to the nearest URRAP road (km) 4,637 44.51 67.92 0.0005 188.89 V. MAIN RESULTS AND DISCUSSION First of all, the fixed-effect panel regression is performed for different types of income sources. When project beneficiaries are defined irrespective of distance from the roads, it is found that only major road improvements have an impact on crop production growth (Table 5). The coefficient of DRSDP is estimated at 0.73, which is statistically significant. On the other hand, the coefficient of DURRAP is negative and insignificant. There are a number of possible reasons for this unexpected result. For instance, the URRAP may actually have had no impact on agricultural income, possibly because of the low technical standards of URRAP roads or the relatively small incremental changes in road condition caused by the URRAP investment. The URRAP roads are normally gravel roads with no significant structures, such as bridges, culverts and drainage. Therefore, sustainability is also a challenge unless systematic maintenance regimes are put in place. This may be another reason for the estimated insignificant effect. However, the more likely reason is that feeder road beneficiaries may be overestimated. Given such low technical standards, it is unlikely that the URRAP roads would benefit those who live far from the project roads. This view is supported by the results shown in the last four columns, which are our preferred models. When beneficiaries are limited to the 2km areas of the project roads, both major and feeder roads are found to have a significant impact. The coefficients of DRSDP2km and DURRAP2km are 0.60 and 0.56, respectively. Both are statistically significant. In addition, the significant interaction term suggests the important synergy between major corridors and feeder roads. The coefficient of DRSDP2km •DURRAP2km is positive and strongly significant.4 This is one of the most important findings in this paper, reconfirming that the road sector is a network: Trunk and feeder roads must be connected 4 The interactive effect between DRSDP and DURRAP cannot be examined because of multicollinearity. - 17 - with one another. This may be an intuitively obvious finding, but there has been little evidence to support this in the literature. One may wonder whether this conclusion is robust regardless of the selection of the threshold (i.e., 2km). When project beneficiaries are defined by an alternative threshold of 5km, the significant impact of DRSDP5km remains unchanged, and the coefficient of DURRAP5km is found to be positive but not significant (Table 6). Thus, the results are consistent: the impact of a major road improvement is sustained over a broader area, but the impact of feeder road improvements is geographically limited. When distance between households and roads is taken into account, this finding can be reconfirmed more clearly. Let DRSDP be replaced with DRSDP•lnKM, and DURRAP with DURRAP•lnKM, respectively. KM represents the distance between households and the nearest project road. The results are shown in the last two columns of the table. The coefficient of DRSDP•lnKM is negative but not significant, meaning that the impact of RSDP roads does not depend on the distance. On the other hand, the coefficient of DURRAP•lnKM is significantly negative at -0.15, meaning that the impact of improved URRAP roads decreases with distance. Numerically, the magnitude of the coefficient implies that the impact would taper off quickly, especially when distance is less than 2km. All the indications are that the impact of major road improvements is broader, while the impact of feeder roads is more area- specific. Hence, the following discussion will be focused on the results based on the 2km distance identification strategy. But all the results are quite robust, as will be shown. Table 5. Fixed-effect panel regression on household crop production income Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. DRSDP 0.734 (0.232) *** 0.735 (0.232) *** DURRAP -0.206 (0.237) -0.207 (0.237) DRSDP2km 0.746 (0.286) *** 0.607 (0.286) ** -0.633 (0.512) DURRAP2km 0.690 (0.277) ** 0.564 (0.278) ** 0.470 (0.281) * DRSDP2km•DURRAP2km 1.472 (0.580) *** lnL 0.049 (0.079) 0.049 (0.079) 0.048 (0.079) 0.050 (0.079) 0.052 (0.079) 0.051 (0.079) 0.049 (0.079) lnR 1.030 (0.072) *** 1.032 (0.073) *** 1.027 (0.072) *** 1.026 (0.073) *** 1.034 (0.073) *** 1.028 (0.073) *** 1.025 (0.073) *** lnI 0.418 (0.095) *** 0.424 (0.095) *** 0.420 (0.095) *** 0.421 (0.095) *** 0.420 (0.095) *** 0.420 (0.095) *** 0.420 (0.095) *** lnP 0.100 (0.042) ** 0.095 (0.043) ** 0.098 (0.042) ** 0.099 (0.042) ** 0.102 (0.043) ** 0.102 (0.042) ** 0.101 (0.042) ** lnF 0.046 (0.043) 0.049 (0.043) 0.049 (0.042) 0.047 (0.043) 0.034 (0.043) 0.037 (0.043) 0.042 (0.043) lnSIZE 0.364 (0.221) * 0.338 (0.223) 0.363 (0.221) * 0.332 (0.222) 0.355 (0.223) 0.346 (0.222) 0.331 (0.222) lnMALE 0.561 (0.294) * 0.570 (0.294) ** 0.568 (0.295) * 0.556 (0.294) * 0.551 (0.294) * 0.548 (0.294) * 0.549 (0.294) * lnAGE 0.606 (0.361) * 0.633 (0.364) * 0.608 (0.362) * 0.614 (0.362) * 0.627 (0.362) * 0.613 (0.361) * 0.599 (0.361) * EDU2 -0.084 (0.325) -0.066 (0.322) -0.080 (0.324) -0.081 (0.325) -0.076 (0.323) -0.085 (0.325) -0.091 (0.326) EDU3 -0.465 (0.304) -0.435 (0.303) -0.450 (0.303) -0.479 (0.303) -0.483 (0.301) * -0.500 (0.302) * -0.501 (0.302) * EDU4 -0.074 (0.648) -0.140 (0.642) -0.051 (0.647) -0.101 (0.646) -0.143 (0.644) -0.096 (0.647) -0.096 (0.648) EDU5 -0.617 (0.454) -0.618 (0.453) -0.597 (0.452) -0.630 (0.452) -0.649 (0.451) -0.640 (0.450) -0.638 (0.450) EDU6 -0.706 (0.592) -0.668 (0.588) -0.700 (0.592) -0.703 (0.585) -0.677 (0.586) -0.701 (0.584) -0.681 (0.583) EDU7 0.012 (0.407) -0.006 (0.408) 0.020 (0.409) 0.004 (0.407) 0.030 (0.403) 0.036 (0.404) 0.042 (0.405) COOK 0.263 (0.453) 0.192 (0.453) 0.264 (0.453) 0.220 (0.453) 0.178 (0.452) 0.204 (0.453) 0.191 (0.454) LIGH -0.680 (0.302) ** -0.694 (0.301) ** -0.684 (0.302) ** -0.688 (0.301) ** -0.633 (0.304) ** -0.641 (0.304) ** -0.632 (0.304) ** T2013 -0.419 (0.187) ** -0.385 (0.190) ** -0.395 (0.189) ** -0.415 (0.187) ** -0.473 (0.190) ** -0.466 (0.191) ** -0.460 (0.191) ** T2014 -0.531 (0.214) ** -0.318 (0.218) -0.455 (0.224) ** -0.467 (0.209) ** -0.576 (0.223) *** -0.602 (0.224) *** -0.590 (0.224) *** T2015 0.652 (0.210) *** 1.047 (0.255) *** 0.795 (0.259) *** 0.754 (0.202) *** 0.663 (0.224) *** 0.584 (0.224) *** 0.587 (0.224) *** T2016 -0.606 (0.217) *** -0.080 (0.249) -0.464 (0.263) * -0.434 (0.203) ** -0.458 (0.214) ** -0.589 (0.220) *** -0.553 (0.221) ** Constant 4.837 (1.629) *** 4.856 (1.635) *** 4.892 (1.626) *** 4.863 (1.631) *** 4.644 (1.635) *** 4.723 (1.630) *** 4.813 (1.628) *** Obs. 4639 4639 4639 4639 4639 4639 4639 - 19 - No. of groups 1004 1004 1004 1004 1004 1004 1004 R-square within 0.2169 0.215 0.217 0.2163 0.2163 0.2172 0.218 between 0.8295 0.8275 0.8307 0.8285 0.8228 0.8244 0.8247 overall 0.6302 0.628 0.631 0.6274 0.6244 0.6245 0.6242 F-stat. 24.05 23.22 23.3 23.37 23.29 22.5 21.68 The dependent variable is the log of total crop production value. Robust standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. Table 6. Fixed-effect panel regression on household crop production income Coef. Std.Err. Coef. Std.Err. DRSDP5km 0.641 (0.275) ** DURRAP5km 0.175 (0.271) DRSDP•KM -0.037 (0.077) DURRAP•KM -0.152 (0.058) *** lnL 0.052 (0.079) 0.050 (0.079) lnR 1.027 (0.073) *** 1.025 (0.072) *** lnI 0.418 (0.095) *** 0.424 (0.095) *** lnP 0.099 (0.042) ** 0.095 (0.043) ** lnF 0.044 (0.043) 0.042 (0.043) lnSIZE 0.336 (0.222) 0.343 (0.222) lnMALE 0.558 (0.294) * 0.579 (0.295) ** lnAGE 0.626 (0.361) * 0.638 (0.363) * EDU2 -0.091 (0.325) -0.069 (0.324) EDU3 -0.485 (0.304) -0.444 (0.302) EDU4 -0.105 (0.646) -0.044 (0.648) EDU5 -0.638 (0.451) -0.593 (0.453) EDU6 -0.695 (0.586) -0.668 (0.588) EDU7 0.019 (0.405) 0.080 (0.409) COOK 0.231 (0.453) 0.203 (0.451) LIGH -0.657 (0.304) ** -0.631 (0.306) ** T2013 -0.433 (0.191) ** -0.409 (0.187) ** T2014 -0.535 (0.225) ** -0.390 (0.207) * T2015 0.664 (0.226) *** 1.074 (0.211) *** T2016 -0.517 (0.220) ** -0.044 (0.204) Constant 4.736 (1.634) *** 4.700 (1.632) *** Obs. 4639 4639 No. of groups 1004 1004 R-square within 0.2163 0.2166 between 0.8282 0.8269 overall 0.6278 0.6255 F-stat. 22.33 23.16 The dependent variable is the log of total crop production value. Robust standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. For other covariates, the estimated results are generally consistent with our prior expectation: The coefficient of L is positive but small and indifferent from zero. Thus, labor is not so productive. This may be attributed partly to our data issue that we assume the whole household members would engage in agricultural activities. But this may not be the case, - 21 - especially when people work on non-agricultural activities. Thus, our labor variable may be overestimated, resulting in underestimating the true impact. The coefficient of rain-fed land, R, is always positive and strongly significant. This is consistent with the fact that Ethiopian agriculture is still land intensive. Advance inputs are generally found to be productive. Particularly, the impacts of irrigation and pesticide use are significant. The coefficient of fertilizer, F, is also positive, though not statistically significant. The findings are consistent with the fact that subsistence farming contributes the majority of agricultural production in Ethiopia. Did the road improvement programs not contribute to improving access to advanced inputs? Equation (3) is estimated using the truncated regression model.5 Proximity to major roads enables farmers to use more fertilizer: the coefficient of DRSDP2km is positive and significant (Table 7). On the other hand, feeder roads do not seem to impact on the farmers’ use of fertilizer. Thus, as far as fertilizer is concerned, the main road network, and not feeder roads, is of particular importance to ensure timely and more affordable availability for farmers. This may be attributed to the fact that Ethiopia imports almost all fertilizer and other advanced inputs through the Port of Djibouti, about 800 km away from the capital city, Addis Ababa. Imported inputs are normally distributed through the main road network to regional cooperatives and farmers. This component of transportation cost contributes a substantial proportion of the fertilizer price to farmers (Rashid et al., 2013). Neither RSDP nor URRAP roads have a significant impact on pesticide use. The coefficient of DRSDP2km is positive but not significant. The coefficient of URRAP is also insignificant. The results are inconclusive, possibly because the use of pesticide is still highly limited in Ethiopia. Recall that pesticides are used by only 10 percent of the surveyed households. 5 It may be open to debate whether the other input variable should be included in the equation. From an empirical point of view, there may be omitted variables that affect both simultaneously, for example, proximity to agrobusinesses, possibly making the result biased. On the other hand, from a practical point of view, there must be complementarities between the different inputs used. Both cases were checked. The results were almost the same. - 22 - For other coefficients, unlike crop production income, the time-specific fixed-effects are estimated to be positive for fertilizer use, meaning that the fertilizer use among farmers systematically increased in comparison with the baseline year, 2012. In addition, the level of education attained by household-heads may also be important. Some education coefficients are significantly positive. This is also a different result from the above crop income regression. Table 7. Truncated regression on input use with fixed-effects Dependent var. lnF lnF lnP lnP Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. DRSDP2km 0.419 (0.205) ** 0.695 (0.220) *** 0.226 (0.499) 0.532 (0.569) DURRAP2km 0.035 (0.151) 0.106 (0.147) -0.369 (0.284) -0.277 (0.281) DRSDP2km•DURRAP2km -0.804 (0.391) ** -1.069 (0.942) lnL 0.020 (0.055) 0.021 (0.055) -0.107 (0.168) -0.108 (0.164) lnR 1.084 (0.079) *** 1.083 (0.080) *** 1.035 (0.165) *** 1.081 (0.163) *** lnI 0.019 (0.027) 0.020 (0.027) 0.052 (0.076) 0.051 (0.075) lnP 0.050 (0.014) *** 0.050 (0.014) *** lnF 0.058 (0.030) ** 0.049 (0.029) * lnSIZE -0.038 (0.204) -0.044 (0.203) -0.652 (0.425) -0.688 (0.425) * lnMALE -0.340 (0.256) -0.345 (0.258) 0.058 (0.300) 0.034 (0.307) lnAGE 0.245 (0.230) 0.243 (0.229) -0.582 (0.484) -0.628 (0.489) EDU2 0.393 (0.273) 0.374 (0.270) -0.569 (0.519) -0.663 (0.524) EDU3 0.491 (0.288) * 0.484 (0.286) * 1.393 (0.627) ** 1.384 (0.633) ** EDU4 0.031 (0.417) 0.220 (0.560) -0.453 (0.377) -0.096 (0.454) EDU5 0.115 (0.310) 0.111 (0.327) -0.296 (0.554) -0.508 (0.610) EDU6 0.936 (0.338) *** 1.099 (0.387) *** -7.335 (0.832) *** -7.337 (0.833) *** EDU7 -0.071 (0.217) -0.056 (0.215) 0.424 (0.316) 0.450 (0.321) COOK 0.379 (0.426) 0.409 (0.429) 0.367 (0.386) 0.393 (0.356) LIGH -0.144 (0.248) -0.185 (0.251) 0.337 (0.509) 0.270 (0.518) T2013 0.157 (0.141) 0.162 (0.141) 0.085 (0.254) 0.116 (0.255) T2014 1.173 (0.179) *** 1.160 (0.177) *** -0.366 (0.229) * -0.378 (0.230) * T2015 0.152 (0.166) 0.139 (0.165) -0.626 (0.221) *** -0.618 (0.221) *** T2016 0.365 (0.170) ** 0.343 (0.169) ** -0.376 (0.306) -0.352 (0.303) Constant -1.471 (1.126) -1.445 (1.125) 4.137 (2.206) * 4.373 (2.254) * Obs. 4639 4639 4639 4639 Of which, truncated 3712 3712 4198 4198 No. of groups 390 390 287 287 Wald chi2 . . . . The dependent variable is the log of fertilizer or pesticide used. Cluster standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. - 23 - As shown above, one likely results chain from road improvements to household income is that improved main roads contributed to farmers’ access to input markets, particularly for fertilizer. How about access to output markets? Equation (5) is estimated using the truncated regression model. Although the improvement of feeder roads was found to have little influence on farmers’ access to input markets, feeder roads do play an important role in providing access to the output market. The coefficient of DURRAP2km is estimated at 0.36 to 0.39, which is significant at least at the 10 percent level (Table 8). By contrast, major roads do not have any impact on farmers’ output sales. This can be interpreted to mean that farmers are selling extra produce at their neighboring markets. This is in line with the above observation that most farmers travel fairly short distances to market (see Figure 3). Therefore, feeder roads are far more critical than main roads, allowing the consolidation of produce which can then by distributed to larger market centers. Table 8. Truncated regression of market crop sales with fixed-effects Dependent var. lnSALE Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. DRSDP2km -0.183 (0.245) -0.244 (0.243) -0.548 (0.415) DURRAP2km 0.366 (0.220) * 0.392 (0.223) * 0.326 (0.236) DRSDP2km•DURRAP2km 0.539 (0.494) lnL 0.040 (0.074) 0.051 (0.076) 0.052 (0.077) 0.048 (0.077) lnR 0.022 (0.036) 0.026 (0.037) 0.028 (0.036) 0.027 (0.037) lnI 0.007 (0.042) 0.003 (0.043) 0.005 (0.043) 0.005 (0.043) lnP -0.009 (0.019) -0.007 (0.020) -0.007 (0.020) -0.008 (0.020) lnF 0.031 (0.019) * 0.024 (0.019) 0.025 (0.019) 0.025 (0.019) lnSIZE -0.090 (0.246) -0.106 (0.246) -0.103 (0.245) -0.095 (0.247) lnMALE 0.240 (0.301) 0.226 (0.293) 0.237 (0.295) 0.217 (0.294) lnAGE 0.241 (0.326) 0.239 (0.322) 0.231 (0.321) 0.234 (0.321) EDU2 0.082 (0.251) 0.079 (0.249) 0.062 (0.250) 0.068 (0.249) EDU3 -0.064 (0.363) -0.060 (0.365) -0.077 (0.365) -0.065 (0.364) EDU4 0.875 (0.504) * 0.936 (0.524) * 0.961 (0.517) * 0.900 (0.523) * EDU5 -0.036 (0.511) -0.007 (0.513) -0.006 (0.510) -0.012 (0.510) EDU6 0.078 (0.578) 0.093 (0.600) 0.137 (0.575) 0.076 (0.598) EDU7 0.070 (0.328) 0.146 (0.334) 0.155 (0.337) 0.136 (0.335) COOK -0.128 (0.467) -0.111 (0.477) -0.120 (0.471) -0.140 (0.467) LIGH 0.390 (0.228) * 0.448 (0.223) ** 0.442 (0.224) * 0.476 (0.220) ** - 24 - T2013 0.537 (0.209) *** 0.507 (0.211) ** 0.502 (0.211) ** 0.503 (0.211) ** T2014 1.018 (0.198) *** 0.904 (0.212) *** 0.909 (0.213) *** 0.927 (0.217) *** T2015 0.134 (0.216) -0.053 (0.240) -0.035 (0.242) -0.020 (0.244) T2016 0.379 (0.214) * 0.176 (0.235) 0.218 (0.237) 0.246 (0.241) Constant 1.928 (1.440) 1.703 (1.428) 1.702 (1.423) 1.710 (1.424) Obs. 4639 4639 4639 4639 Of which, truncated 3595 3595 3595 3595 No. of groups 444 444 444 444 Wald chi2 . . . . The dependent variable is the log of crop sales at market. Cluster standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. Turning to other types of household income, no impact is found on livestock (Table 9). This may be able to be understood because pastoralism has a certain mobile aspect and is less transport-intensive. In Ethiopia, many pastoralists are scattered in remote areas. Thus, the quality of transport infrastructure may not matter much, though livestock can be reared using intensive methods in non-pastoral areas and obtaining feed and forage for the stock and delivering the stock to slaughter and urban markets is potentially more dependent on good trunk roads. On the other hand, non-agricultural income seems be affected positively by improved RSDP roads and negatively by URRAP roads. This can be interpreted to mean that trunk road improvements generated some indirect impacts in the project areas. Given the increase in crop production and market sales, local economies may be boosted by increased activity in non-agricultural sectors. Interestingly, this impact seems to have intensified over time: the time-specific fixed-effects are significant and positive. The coefficients of T become larger the more time passes after the road improvement. The current household survey data do not allow the possible reasons behind this non-agricultural growth to be identified. Some additional data will now be examined. - 25 - Table 9. Fixed-effect panel regression on livestock and nonagricultural income Livestock income Livestock income NonAG income NonAG income Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. DRSDP2km -0.638 (0.304) ** 0.564 (0.997) 2.018 (0.396) *** -1.105 (0.982) DURRAP2km 0.119 (0.325) 0.206 (0.325) -1.414 (0.379) *** -1.640 (0.383) *** DRSDP2km•DURRAP2km -1.425 (1.010) 3.701 (1.042) *** lnL 0.166 (0.091) * 0.168 (0.091) * 0.160 (0.109) 0.155 (0.109) lnG 0.056 (0.094) 0.056 (0.094) lnSIZE 0.332 (0.265) 0.347 (0.264) 0.421 (0.320) 0.383 (0.319) lnMALE 0.472 (0.353) 0.471 (0.352) -0.475 (0.417) -0.472 (0.415) lnAGE 0.684 (0.464) 0.698 (0.464) -0.646 (0.502) -0.681 (0.501) EDU2 -0.175 (0.364) -0.169 (0.364) -0.087 (0.430) -0.102 (0.430) EDU3 -0.612 (0.369) * -0.613 (0.369) * 0.919 (0.455) ** 0.919 (0.453) ** EDU4 -0.440 (0.683) -0.440 (0.682) 1.463 (0.699) ** 1.463 (0.697) ** EDU5 -0.973 (0.398) ** -0.975 (0.401) ** 0.507 (0.560) 0.512 (0.553) EDU6 -1.648 (0.546) *** -1.667 (0.545) *** 1.339 (0.757) * 1.388 (0.748) * EDU7 0.000 (0.533) -0.006 (0.533) 0.181 (0.542) 0.196 (0.543) COOK -0.537 (0.403) -0.523 (0.403) 0.355 (0.445) 0.319 (0.443) LIGH -0.597 (0.331) * -0.607 (0.331) * 1.597 (0.422) *** 1.621 (0.420) *** T2013 2.134 (0.240) *** 2.125 (0.239) *** 4.909 (0.296) *** 4.933 (0.297) *** T2014 2.270 (0.270) *** 2.255 (0.269) *** 4.649 (0.308) *** 4.688 (0.308) *** T2015 2.109 (0.280) *** 2.104 (0.280) *** 6.529 (0.323) *** 6.541 (0.324) *** T2016 2.140 (0.305) *** 2.102 (0.305) *** 6.110 (0.335) *** 6.210 (0.338) *** Constant -7.580 (1.867) *** -7.652 (1.865) *** 0.788 (1.920) 0.986 (1.917) Obs. 4639 4639 4639 4639 No. of groups 1004 1004 1004 1004 R-square within 0.0406 0.0412 0.1861 0.189 between 0.2514 0.2507 0.2791 0.2785 overall 0.1145 0.1164 0.1912 0.1949 F-stat. 8.8 8.39 38.2 37.32 The dependent variable is the log of livestock or nonagricultural income. Robust standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. To investigate the indirect effects of improved road accessibility more deeply, two additional pieces of data are used, which were collected together with the household surveys. First, the demand for transport services is estimated from passenger survey data. Note that the surveyed passengers may not be the same persons over time, although the surveys were carried out at the same locations that are normally village centers where formal and informal transport services are offered. Thus, the data are cross-sectional at five periods of time and reflect overall demand for transport. - 26 - The sample size is much smaller than the above household data: 30 passengers were surveyed along each RSDP road every year. While the total sample size is 600 over five years, the following analysis uses only 498 observations because of some missing values. Each passenger was asked about their travel time, frequency, costs and mode, as well as some basic individual characteristics, such as age and education attained (Table 10). In the sample, about half of the surveyed passengers travel twice to four times per month. About 35 percent travel only once a month. Table 10. Summary statistics of passenger survey data Variable Abb. Obs. Mean Std.Dev. Min Max Frequency of travel per month Once 172 2 to 4 times 243 5 to 8 times 83 Dummy variable for program beneficiaries DRSDP 498 0.227 0.419 0 1 from RSDP roads Distance to all weather road (minutes): 15 to 30 minutes FEED15-30m 498 0.235 0.424 0 1 31 to 60 minutes FEED30-60m 498 0.098 0.298 0 1 1 to 2 hours FEED1-2hr 498 0.056 0.231 0 1 2 to 3 hours FEED2-3hr 498 0.044 0.206 0 1 More than 3 hours FEED>3hr 498 0.052 0.223 0 1 Average travel time (minutes) TIME 498 103.886 128.402 0 850 Average transport fares (EHB) FARE 498 36.040 45.574 0 551 Average waiting time before boarding WAIT 498 77.966 84.962 0 480 (minutes) Dummy variable for occupation: Farmers FARM 498 0.112 0.316 0 1 Traders, bakery and other small businesses SME 498 0.173 0.378 0 1 Government employees, nurse, teachers GOVT 498 0.205 0.404 0 1 Students STUD 498 0.122 0.328 0 1 Dummy variable for transport mode: Minibus MMinibus 498 0.514 0.500 0 1 Bajaj, motorcycles MBajaj, Motor 498 0.008 0.089 0 1 Age AGE 498 33.558 11.717 14 87 Dummy variable for male MALE 498 0.624 0.485 0 1 Dummy variable for education attained: Higher 2EDU 498 0.390 0.488 0 1 than secondary school Dummy variable for living areas: Urban URBAN 498 0.729 0.445 0 1 Time fixed effects: t=2013 T2013 498 0.175 0.380 0 1 - 27 - t=2014 T2014 498 0.175 0.380 0 1 t=2015 T2015 498 0.237 0.426 0 1 t=2016 T2016 498 0.229 0.421 0 1 The censored regression is performed where the dependent variable is the number of journeys per month. Our dependent variable is censored to (i) once, (ii) 2 to 4 times, (iii) 5 to 8 times, and (iv) more than 8 times, because of our simplified survey design.6 The results show that the mobility of people is crucially dependent on proximity to the nearest feeder road in good condition (Table 11). It is significantly reduced when people live more than 1 hour from an all-weather road. The coefficients of FEED1-2hr and FEED2-3hr are estimated to be significantly negative. On the other hand, the proximity to RSDP roads does not seem to affect people’s demand for transportation. To examine the interaction effect between RSDP and feeder roads, a dummy variable representing those who have a good feeder road in less than 1 hour is created. As expected, this variable has a positive and significant coefficient of 0.563. However, there is no significant interaction effect on the transport demand between the main and feeder road accessibility. The estimated demand equation also indicates what constraints on mobility people face in rural Ethiopia. One of them is efficiency of transport services: The coefficient of travel time, TIME, is significantly negative, meaning transport speed is important to meet people’s demand for transportation. The waiting time variable (WAIT) also has a negative and weakly significant coefficient. That is, transportation demand would be reduced when waiting time is longer. The finding is in line with the positive coefficient of bajaj, which is a three-wheeler taxi now widely used in Ethiopia. Frequency of transport services really matters. This is why people seem to prefer to use bajajs, because they do not have to wait for a taxi or bus to be fully laden with passengers before departing. On the other hand, transport fares have a negative coefficient, as expected, however, there is no statistical significance. 6 In our sample, no one responded that she traveled more than 8 times. Thus, we have only 3 discrete choices that are actually observed. - 28 - The transport service demand is found to be significantly high when people engage in retail and other local businesses, such as bakery and carpentry: The coefficient of SME is significant. Farmers’ demand seems to be weak: the coefficient of FARM is small and statistically insignificant. This is broadly consistent with the above. The estimation results suggest that there is an important indirect effect on local businesses. As the feeder road network improves, agricultural production and market sales increase. Together with such agricultural growth, local economies will grow, with more transport services provided. This is a much broader impact that improved roads can generate. Table 11. Censored regression for transportation demand Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. RSDP D -0.265 (0.223) -0.161 (0.214) 0.066 (0.339) FEED15-30min 0.192 (0.210) FEED30-60min -0.323 (0.292) FEED1-2hr -0.611 (0.282) ** FEED2-3hr -1.131 (0.327) *** FEED>3hr -0.086 (0.364) DLess than 1hr 0.563 (0.209) *** 0.642 (0.231) *** RSDP Less than 1hr D •D -0.280 (0.390) TIME -0.0018 (0.0007) ** -0.0018 (0.0007) ** -0.0018 (0.0007) ** FARE -0.0041 (0.0026) -0.0041 (0.0025) -0.0040 (0.0025) WAIT -0.0014 (0.0008) * -0.0010 (0.0008) -0.0010 (0.0008) FARM 0.029 (0.250) 0.043 (0.244) 0.025 (0.246) SME 0.941 (0.245) *** 0.926 (0.246) *** 0.922 (0.247) *** GOVT 0.022 (0.196) 0.002 (0.197) -0.010 (0.197) STUD 0.422 (0.246) * 0.409 (0.251) * 0.410 (0.252) * MMinibus -0.187 (0.177) -0.135 (0.172) -0.140 (0.173) MBajaj, Motor 0.812 (0.528) 0.924 (0.558) * 0.946 (0.556) * AGE -0.004 (0.007) -0.003 (0.007) -0.003 (0.007) MALE 0.262 (0.161) * 0.207 (0.162) 0.211 (0.162) 2EDU -0.033 (0.160) -0.026 (0.159) -0.024 (0.159) URBAN -0.046 (0.211) 0.056 (0.196) 0.049 (0.196) T2013 -0.119 (0.250) -0.028 (0.246) -0.026 (0.246) T2014 0.828 (0.271) *** 0.809 (0.274) *** 0.811 (0.275) *** T2015 -0.036 (0.246) -0.012 (0.246) -0.027 (0.245) T2016 0.360 (0.292) 0.346 (0.295) 0.355 (0.298) Constant 3.037 (0.419) *** 2.289 (0.402) *** 2.231 (0.412) *** Obs. 498 498 498 Wald chi2 102.05 91.53 91.98 - 29 - The dependent variable is the (censored) frequency of travel per month. Robust standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. Another set of complementary data covers local businesses. Along with the household surveys, local enterprises were also interviewed. Again, the sample size is limited; Only 30 firms, such as retailers and restaurants, were surveyed along each RSDP road every year. Each firm was asked about the number of employees and their basic business environment. In the sample, the average size of firms is about 5 employees (Table 12). Logarithmic panel regression was performed on the size of firms. It was found that the improvement of RSDP roads had a positive and significant impact on the firm size (Table 13). This is consistent with the above local business hypothesis: as the local economies pick up, local businesses are hiring more people. The demand for local services increases along the main corridors. Feeder road accessibility is also important: The size of firms is significantly reduced when they are located more than 30 minutes from an all-weather road. The coefficients of FEED30-60min and FEED>3hr are significantly negative.7 Though, there is no clear synergy between the RSDP and feeder road proximity. By type of business, hotels, restaurants and pharmacies are among the rapidly growing subsectors, compared with a baseline, which is retailers. As the traffic increases, these businesses along the improved roads are likely to expand, generating more local jobs. While supply increases, prices are also likely to decline. These indirect effects are considered as a main reason why the road programs have a positive and significant impact on non- agricultural household income. 7 Note that in our sample data, there are no firms that are located 1 to 2 hours from an all weather road. - 30 - Table 12. Summary statistics of local business survey data Variable Abb. Obs. Mean Std.Dev. Min Max Number of employees 474 4.989 6.139 1 53 Dummy variable for program DRSDP beneficiaries from RSDP roads 474 0.234 0.424 0 1 Distance to all weather road (minutes): 15 to 30 minutes FEED15-30min 474 0.139 0.347 0 1 31 to 60 minutes FEED30-60min 474 0.034 0.181 0 1 1 to 2 hours 474 0 0 0 0 2 to 3 hours FEED2-3hr 474 0.053 0.224 0 1 More than 3 hours FEED>3hr 474 0.015 0.121 0 1 Business perspective compared to last year sales: No change NOCH 474 0.095 0.293 0 1 Increasing INCRE 474 0.859 0.349 0 1 Business type: Shop 474 0.367 0.483 0 1 Hotel, restaurant, café HOTL 474 0.340 0.474 0 1 Pharmacy PHAM 474 0.293 0.456 0 1 Time fixed effects: t=2013 T2013 474 0.224 0.417 0 1 t=2014 T2014 474 0.184 0.388 0 1 t=2015 T2015 474 0.184 0.388 0 1 t=2016 T2016 474 0.173 0.379 0 1 Table 13. OLS regression on size of firms Coef. Std.Err. Coef. Std.Err. Coef. Std.Err. RSDP D 0.310 (0.074) *** 0.304 (0.077) *** 0.197 (0.142) FEED15-30min -0.056 (0.077) FEED30-60min -0.499 (0.155) *** FEED2-3hr -0.026 (0.113) FEED>3hr -0.468 (0.112) *** DLess than 30min 0.244 (0.089) *** 0.227 (0.102) ** RSDP Less than 30min D •D 0.116 (0.151) NOCH 0.067 (0.126) 0.057 (0.123) 0.060 (0.124) INCRE 0.191 (0.119) * 0.191 (0.117) * 0.194 (0.118) * HOTL 1.062 (0.071) *** 1.059 (0.071) *** 1.060 (0.071) *** PHAM 0.336 (0.046) *** 0.346 (0.046) *** 0.347 (0.046) *** T2013 0.254 (0.087) *** 0.251 (0.085) *** 0.251 (0.085) *** T2014 0.101 (0.089) 0.097 (0.092) 0.092 (0.093) T2015 0.047 (0.092) 0.044 (0.096) 0.039 (0.097) - 31 - T2016 0.237 (0.102) ** 0.255 (0.103) ** 0.253 (0.103) ** Constant 0.469 (0.125) *** 0.216 (0.158) 0.230 (0.163) Obs. 474 474 474 R-squared 0.440 0.430 0.430 F-stat. 25.69 30.77 27.85 The dependent variable is the log of the number of employees. Robust standard errors are shown in parentheses. *, ** and *** indicate the statistical significance at the 10, 5 and 1 percent, respectively. VI. CONCLUSIONS Rural access is among the most important infrastructure constraints in rural Africa. Despite a relatively large number of earlier studies in this area, there are still divergent views on the results chains generated by rural access improvement. The paper examines the heterogeneous impacts of road accessibility on agriculture and non-agricultural growth. It is found that it is important to distinguish between different types of road to understand these heterogeneous effects. Using the comprehensive household and other data from Ethiopia, the paper examined the impacts of corridor improvement and feeder road rehabilitation. It was found that crop production is increased by both major and feeder road accessibility. Significant synergy between them was also found. When investigating further into this effect, there are actually two different impacts: Farmers’ access to input markets, especially fertilizer, was improved mainly by major corridor improvement. On the other hand, output market access was improved by feeder road improvement, though this impact seems to be relatively weak. These findings may make sense given the current Ethiopian context. The timely distribution of fertilizer relies on the main road network, because many agricultural inputs are imported from abroad. Farmers may sell extra produce in their neighboring markets, but the market participation is generally still limited. It is also found that there is no significant impact of accessibility on household income from livestock. By contrast, the households’ non-agricultural income is increased by improved - 32 - road connectivity. There seem to be other indirect effects. With additional data, further investigation was carried out: It is found that along the improved roads, significant transport demand was generated by local business activity. As the road condition was improved, the local economies grew with more local jobs created. This is the main source of the secondary effects of road improvements. To meet the increasing demand for transport services, efficiency and frequency are found to be important. The elasticities are small but statistically significant. By contrast, transport costs do not seem to be critical. Thus, not only physical infrastructure, regardless of whether it is a main road or a feeder road, but also transport services are also important to promote economic growth in rural Ethiopia. - 33 - REFERENCES Bell, Clive and Susanne van Dillen. 2014. How does India’s rural roads program affect the grassroots? Findings from a survey in Upland Orissa. Land Economics, Vol. 90(2), pp. 372-394. Boudier-Bensebaa, Fabienne. 2005. “Agglomeration Economies and Location Choice: Foreign Direct Investment in Hungary.” Economics of Transition 13, no. 4: 605-628. Bravo-Ortega, and Lederman. 2004. Agricultural productivity and its determinants: Revisiting international experiences. Estudios de Economia, Vol. 31(2), pp. 133-163. Cieślik, Andrzej, and Michael Ryan. 2004. Explaining Japanese Direct Investment Flows into an Enlarged Europe: A Comparison of Gravity and Economic Potential Approaches. Journal of the Japanese and International Economies 8 (1): 12–37. Damania, Richard, Claudia Berg, Jason Russ, Federico Barra, John Nash, and Rubaba Ali. 2017. Agricultural technology choice and transport. American Journal of Agricultural Economics, Vol. 99(1), pp. 265–284 Danida. 2010. “Impact Evaluation of Danida Support to Rural Transport Infrastructure in Nicaragua.” Evaluation Department, Ministry of Foreign Affairs of Denmark Datta, Saugato. 2012. The impact of improved highways on Indian firms. Journal of Development Economics, Vol. 99(1), pp. 46-57. Dercon, Stefan, Daniel Gilligan, John Hoddinott and Tassew Woldehanna. 2009. The impact of agricultural extension and roads on poverty and consumption growth in fifteen Ethiopian villages, American Journal of Agricultural Economics, Vol. 91(4), pp. 1007- 1021. Dercon, Stefan, John Hoddinott, and Tassew Woldehanna. 2012. Growth and chronic poverty: Evidence from rural communities in Ethiopian. Journal of Development Studies, Vol. 48(2), pp. 238-253. Diao, Xinshen, James Thurlow, Samuel Benin, and Shenggen Fan. 2012. Strategies and Priorities for African Agriculture Economywide Perspectives from Country Studies. International Food Policy Research Institute. Dorosh, Wang, You, and Schmidt. 2012. Road connectivity, population, and crop production in Sub-Saharan Africa. Agricultural Economics, Vol. 43, pp. 89-103 Dorosh, Paul, and James Thurlow. 2017. Beyond agriculture versus non-agriculture: Decomposing sectoral growth-poverty linkages in five African countries. World Development, forthcoming. Escobal, Javier, and Carmen Ponce. 2002. The benefits of rural roads: Enhancing income opportunities for the rural poor. GRADE, Working Paper 40–1, Lima. Gyimah-Brempong. 1987. Scale elasticities in Ghanaian cocoa production. Applied Economics, Vol. 19, pp. 1383-1390. - 34 - Jalan, J. and Martin Ravallion. 1998. “Are There Dynamic Gains from a Poor-area Development Program?” Journal of Public Economics 67(1): 65–85. Key, Nigel, Elisabeth Sadoulet, and de Janvry, Alain, 2000. Transactions costs and agricultural household supply response. American Journal of Agricultural Economics, Vol. 82(2), pp. 245–59. Khandker, Shahidur, Zaid Bakht, Gayatri Koolwal. 2009. The poverty impact of rural roads: Evidence from Bangladesh. Economic Development and Cultural Change, Vol. 57(4), pp. 685-722. Khandker, Shahidur, Gayatri Koolwal. 2011. Estimating the long-term impacts of rural roads: a dynamic panel approach. Policy Research Working Paper No. 5867, World Bank. Lokshin, Michael, and Luslan Yemtsov. 2005. Has rural infrastructure rehabilitation in Georgia helped the poor? The World Bank Economic Review, Vol. 19(2), pp. 311-333. Mu, Ren, and Dominique van de Walle. 2011. Rural roads and local market development in Vietnam. Journal of Development Studies, Vol. 47(5), pp. 709-734. Nakamura, Shohei. 2017. Review of Ethiopia rural road program: Welfare impacts of recent rural road development in Ethiopia. Unpublished. Rand, John. 2011. Evaluating the employment-generating impact of rural roads in Nicaragua. Journal of Development Effectiveness, Vol. 3(1), pp. 28-43. Rashid, Shahidur, Nigussie Tefera, Nicholas Minot, Gezahengn Ayele. 2013. Fertilizer in Ethiopia. IFPRI Discussion Paper 01304, International Food Policy Research Institute. Sheahan, Megan, and Christopher Barrett. 2017. Ten striking facts about agricultural input use in Sub-Saharan Africa. Food Policy, Vol. 67, pp. 12-25. Shirley, Chad, and Clifford Winston. 2004. “Firm Inventory Behavior and the Returns from Highway Infrastructure Investment.” Journal of Urban Economics 55, no. 2: 398-415. Spielman, David, Dawit Mekonnen, and Dawit Alemu, 2012. Seed, fertilizer, and agricultural extension in Ethiopia, IFPRI book chapters, in: Food and agriculture in Ethiopia: Progress and policy challenges, chapter 4 International Food Policy Research Institute. Taffesse, Alemayehu Seyoum; Dorosh, Paul A. and Gemessa, Sinafikeh Asrat. 2012. Crop production in Ethiopia: Regional patterns and trends. In Food and Agriculture in Ethiopia: Progress and Policy Challenges, ed. Paul A. Dorosh and Shahidur Rashid. Chapter 3 pp. 53-83. Philadelphia, PA: University of Pennsylvania Press. Warr, Peter. 2008. How road improvement reduces poverty: The case of Laos. Agricultural Economics, Vol. 39(3), pp. 269-279.