WPS 3875 The Poverty Impact of Rural Roads: Evidence from Bangladesh1 Shahidur R. Khandker World Bank Zaid Bakht Bangladesh Institute of Development Studies Gayatri B. Koolwal National Economic Research Associates Abstract The rationale for public investment in rural roads is that households can better exploit agricultural and non-agricultural opportunities to employ labor and capital more efficiently. But significant knowledge gaps remain as to how opportunities provided by roads actually filter back into household outcomes and their distributional consequences. This paper examines the impacts of rural road projects using household-level panel data from Bangladesh. Rural road investments are found to reduce poverty significantly through higher agricultural production, higher wages, lower input and transportation costs, and higher output prices. Rural roads also lead to higher girls' and boys' schooling. Road investments are pro-poor, meaning the gains are proportionately higher for the poor than for the non-poor. World Bank Policy Research Working Paper 3875, April 2006 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 view of the World Bank, its Executive Directors, or the countries they represent. Policy Research Working Papers are available online at http://econ.worldbank.org. 1We would like to thank Gershon Feder and Robert Floyd for valuable comments and suggestions, as well as Hussain Samad for research assistance. 1 The Poverty Impact of Rural Roads: Evidence from Bangladesh 1. Introduction Improved roads and infrastructure can create opportunities for economic growth and poverty reduction through a range of mechanisms. Roads reduce transportation costs and the costs of consumption and production of goods and services (BIDS 2004). With easier access to markets and technology, improved roads expand farm and nonfarm production through increased availability of relevant inputs and lower input costs (Binswanger, Khandker, and Rosenzweig 1993; BIDS 2004; Levy 1996) as well as growth in rural enterprises (Lokshin and Yemtsov 2005). At the household level, road development contributes to higher productivity and demand for labor (Leinbach 1983; World Bank 2000), and improved education and health, including for women and girls (Bryceson and Howe 1993; Levy 1996). Road-related studies have also suggested that household consumption is likely to get a boost from increased household income, consequently reducing poverty (BIDS 2004; Fan, Hazell, and Thorat 2000). It has been difficult to precisely quantify the benefits of roads, due to methodological constraints and data limitations. While transport investments have consistently represented 15-20 percent of the World Bank's lending portfolio2, traditional estimates of the returns to such investments using the internal rate of return approach are generally so low that the investments do not appear viable. Such approaches have also been criticized for not being able to capture the true distributional benefits to the targeted population, particularly for the poor (van de Walle, 2002). Additional mechanisms also depend on the type of road improvement project.3 These problems are compounded by the fact that the effects of rural roads are also long-term and thus cannot be captured through the use of cross-sectional data, particularly since unobserved fixed area characteristics influence the placement of road investment in a village or community (Binswanger, Khandker and Rosenzweig, 1993). Since the timeline for the full impact of road improvements to take effect is extremely long, panel data as well as careful selection of control areas are essential in examining the exact impact of road 2 Indeed, transport expenditure shares have remained fairly constant in this range between 1999 and 2004. Within South Asia, this share jumps to 26% of the Bank's lending in that region, as compared to 27% for East Asia and the Pacific, 21% for African countries, 14% for Latin America and the Caribbean, 7% for Middle East and North Africa, and 5% for Europe and Central Asia. 3Such projects, for example, can range from paving dirt roads to facilitation of two-way traffic, straightening, or upgrade to dual carriageway or motorway status. 2 development.4 A few recent studies have used improved scientific approaches for data collection to estimate the impact of roads. Lokshin and Yemtsov (2005) demonstrate the economic benefits of roads and other infrastructural projects, using a propensity score-matched double difference method. van de Walle and Cratty (2005) also use the same method to estimate the impact of road development.5 Finally, BIDS (2004) employs a panel household survey data of a quasi-experimental nature to assess the impacts of road improvement projects. However, the BIDS (2004) study employs bivariate analysis (e.g., a difference-in-difference technique), which is not ideal for this dataset given that the data are not of random experimentation. Examining the distributional consequences of road development across different income levels has been an additional challenge, particularly on top of the need for a methodology that estimates effects over time and is able to control for unobserved heterogeneity. Aside from Lokshin and Yemtsov (2005), who also conduct their analysis on "poor" and "non-poor" samples to find that each group benefits differently from road development,6 distributional effects of rural road investments have not been addressed extensively in the rural road literature, hampering assessments of the true breakdown of poverty reduction from infrastructural improvements. Our paper, using the BIDS panel data, estimates the income-consumption benefits of road investment by controlling for both household- and community-level heterogeneity. Because poverty reduction is an overarching goal of policymakers and donors, we also assess the poverty reduction effect of road investment projects. We offer an empirical assessment on the impact and role of roads on economic behavior across different income groups by using a fixed-effects quantile estimation approach. Finally, we estimate the distribution of road benefits by gender, examining whether road investment benefits men and women as well as boys and girls differently. The paper is organized as follows. Section 2 presents an econometric framework to estimate the impact of road development. Section 3 discusses the panel data, spanning two project samples, and their respective control villages. Section 4 discusses the results and presents the estimates of economic rates of return to road investment. Section 5 discusses the distribution of road benefits accrued by households across different percentiles of per capita consumption. Section 6 shows how the effect of road 4In previous studies, control areas have often been selected retroactively, contaminating the data selection with the econometrician's knowledge of actual outcomes, as well as reducing much of the data to less-than-reliable, retrospective information. 5The main intent of this paper, however, is to assess the fungibility of development aid between ex ante targeted and non-targeted areas, presenting road improvement as one type of aid initiative. 6They find that the non-poor benefited more from roads through improved medical emergency assistance and opportunities for non-agricultural employment; in contrast, the poor benefited more from female off-farm employment opportunities. 3 development on consumption can be translated into poverty impacts, and presents the poverty effects of road investment across different types of households. The concluding section summarizes the results. 2. Assessing benefits of road investment: An econometric framework Approaches to road evaluation can be distinguished by how wide a net they cast in the search for road impacts. Traditional cost-benefit evaluation of roads has focused on the measurable outputs of road improvements, namely road length, number of beneficiaries, reduced travel times, accident risk, transportation costs, and direct environmental consequences.7 These methods are often based on the assumption that lower agricultural input costs brought about by a road project lead to an increased demand for transport, and hence increased farm production. However, the effects of road improvements spring more generally from improved access to, and functioning of, markets (especially for key production inputs and outputs) as well as key facilities such as education and healthcare. Where existing traffic is small or negligible, estimates of transport demand are only credible if a detailed analysis of the production system is carried out. The broadest evaluation would also look at induced changes in the household production and consumption mix as well as social and political participation, including diversification of income sources, changes in capital accumulation patterns, and interaction with government policy. Binswanger, Khandker, and Rosenzweig (1993) show how roads and other infrastructural investments affect the relationships between input and output markets, household income and employment, and government policy interventions; these in turn are shown to be governed jointly by agroclimatic endowments and agricultural opportunities. Similarly, van de Walle (2002) proposes a diversified, operational approach to measuring the benefits of rural roads, where social welfare from a road is maximized with respect to the total cost of all proposed links, with reliance on community authorities and local residents in project appraisal.8 Our paper uses an econometric approach to estimate the impacts of road investment on household poverty and other rural household outcomes. We allow these outcomes to be directly influenced by agroclimatic and other community endowments as well as agricultural opportunities of a village/community. These observed and unobserved characteristics affect input and output markets, public investments in infrastructure such as roads, and government pricing, interest, and public spending 7The choice among different types of infrastructural programs has traditionally been based on cost-benefit analysis; however, this approach often ignores the externalities that infrastructure programs also generate. In a study of transport in Bangladesh, for example, the benefit-cost ratio of paving a road was calculated as 1.19 based upon existing traffic, but increased to 3.48 when projected increases in traffic were accounted for (Ahmed and Hossain, 1990). 8Social welfare in van de Walle's study is a weighted sum of the average user's social equity value in a community targeted by a particular road link, multiplied by the per-person efficiency gain and the number of people in that community. 4 decisions. Household outcomes are influenced by input and output markets, infrastructure, and government policy. Likewise, infrastructural investments also affect these input and output markets. Finally, government policymaking in credit and other markets (for crops such as paddy, for example) can also influence household outcomes directly and indirectly via the input and output markets as well as infrastructural investments. Such complex interactions make it difficult to identify the precise role of rural road investment on income, productivity, poverty, and human capital investment. The following semi-logarithmic reduced-form income equation, conditional on road investment, can be written as lnYij = Hij + Vj + R + + + ij y (1) j j i where Yij is the per capita income or consumption of i-th household living in j-th village, H is set of observed household characteristics, R is an indicator of the road development project at the village level, V represents observed non-road village-level characteristics, is unobserved village-specific heterogeneity, represents unobserved household characteristics, and is a vector of idiosyncratic errors distributed across households. Similar equations can be written for other outcomes such as prices (P) and institutional infrastructure (I) that are impacted by roads. Since income or consumption is also a function of input and output prices (P) as well as institutional infrastructure (I), road investment has a direct effect on household consumption as well as an indirect effect through prices and institutions. Thus, the total effect of road investment might be decomposed as: K L d lnYij / dRj = lnYij / Rj + (lnYijk / Pjk )(Pjk / Rj ) + ( lnYij / I )(I / Rj ) jl jl k=1 l=1 where dY / dR is the total derivative of the effect of road investment on outcomes, Y / R is the partial effect of road investment, for example, given prices and other factors, Y / P is the partial effect of prices and other intervening factors on outcomes of interest, and P / R is the partial effect of local road conditions on price and similar intervening factors. Similarly, one can obtain partial effects through changes in the institutional factors (I). The changes are measured at i-th household in village j, for the k- th type of intervening price factors, and l-th type of institutional factors. As the road intervention is a community-level variable, consider the community impact of road investment. In this case, one would assemble a sample of communities with varying levels of road improvements and regress the welfare outcome of interest on the road improvement variable. Several factors complicate this exercise, however. The most important concern is that community-level characteristics that are often unobserved to the researcher, such as agro-climatic endowments and 5 agricultural opportunities, may affect both the placement of the road improvement and the welfare outcome of interest. The standard solution for this unobserved heterogeneity bias is to assume that it is not time- varying and, therefore, can be controlled in a panel data regression with fixed community effects. Clearly, if this strategy is to work, there must be some variation in road status over time in the sample communities. The simplest way to proceed is if the data allow a clean division between periods when the road program is in effect (t=1) and when it is not (t=0), where the program villages are selected randomly. This allows a straightforward before-and-after comparison of welfare outcomes between the program and non-program villages and is the basis of the well-known difference-in-difference estimate, where the effect of the policy is estimated by the difference in the relevant outcome for program and non-program areas across the two periods. Such estimates usually are not possible, however, because the decomposition into pre- and post- treatment years is often not available and/or the villages are not selected at random. In the absence of random selection, a difference-in-difference estimate can be constructed with a set of controls for other factors, including time. Also, community-level fixed-effects cannot resolve the bias if household unobserved heterogeneity influences how individual households accrue benefits from road investment. In this case, using household-level fixed-effects rather than community fixed-effects is the appropriate solution. Thus, a household-level panel would be required to resolve both the household and community heterogeneity that affect the estimates of the road investment. This can be done after introducing time- variation in the outcomes and explanatory variables including time in the estimation. Consider the following revised equation of (1): lnYijt = Hijt + Vjt +Rjt + + +ijt y (2) j i Taking the difference over the two-year period of study, one would obtain the following difference equation, where the sources of endogeneity (i.e. the unobserved village and household characteristics, and assuming that these characteristics do not change over time) are dropped out. In this case, the simple OLS can be applied to the following differenced equation to estimate unbiased effect of road development ( ): lnYij =Hij + Vj +Rj + ij (3) 6 3. Data characteristics The panel data used in this paper, collected by the Bangladesh Institute of Development Studies (BIDS), are based on household and community surveys prior to, and following, implementation of two World Bank funded projects that allowed identification of control and treatment villages. The data are used here to calculate economic returns to roads and its impact on poverty, and overcome some of the pitfalls of past road evaluations that have relied mostly on cross-sectional household survey data. The datasets have a true before/after and with/without structure and are reasonably large, allowing a study of household-level impacts, especially with reference to households above and below the poverty line. They cover not just the standard road project outputs such as trip frequency and duration, but also key outcomes (consumption, employment) and a broad range of market interactions. The data collection was financed under the World Bank-funded projects and conducted by BIDS as part of the government's efforts to analyze and quantify both the short-term and long-term impacts of rural road improvements. The first survey covered the Rural Roads and Markets Improvement and Maintenance Project ­ I (RRMIMP-I), which was a component of the Rural Development Project-7 (RDP-7). For the purposes of this paper, we will define this project as RDP. The RDP initiative entailed improvement of 47 FRBs to bitumen surfaced standard, upgrading of 65 secondary markets and construction of 3700 meters of culverts and small bridges. Its physical works were completed during 1995-96. The first phase of the survey collected benchmark information in the study area prior to the project work, and the second phase collected the same information during the first half of 2000. The next phase included 24 project and 18 control villages, and 1,260 households. The second survey studied the Rural Roads and Markets Improvement and Maintenance Project ­ II (RRMIMP-II), which included improvement of 574 kilometers of FRBs to bitumen-surfaced standard, construction of 1,900 meters of culverts, 1,750 meters of bridges and 2,200 meters of small drainage structures on rural roads, and improvement of 136 Growth Center Markets and 41 Ghats (river jetties for boats and vessels). The first phase of the RRMIMP survey collected benchmark information on 872 households from 18 villages during May-September 1997, and the second phase covered the same households between August, 2000-February, 2001. Both projects were funded by the World Bank as part of its effort to promote rural infrastructural development, and, consequently, rural growth and poverty reduction. In both project and control road areas, one roadside village was selected for each road project. The selected roadside villages indicate the immediate influence areas of road intervention. In addition, one remote village was selected for each of the sample roads to assess the decay effects of road development. The remote villages were similar to the roadside villages and were chosen such that they 7 were at least 2 kilometers away from the study or any other paved road. About 50 households were selected from each study village using a stratified random sampling procedure. Both surveys collected a variety of information on general household characteristics, education, healthcare treatment, wage and self-employment, credit activities, assets, income, consumption, marriage, and fertility. Additionally, information on community characteristics and transportation was also collected.9 The outcomes of interest include variables such as household transport expenses, fertilizer price, male agricultural wage, agricultural output and price indices, and household outcomes such as per capita expenditure, male and female labor supply, and boys' and girls' schooling. Road improvements affect the household through changes in three mechanisms: (1) transportation costs as well as input and output prices; (2) labor supply, as well as farm and non-farm production; and (3) household outcomes such as earnings, consumption, and schooling. In our results, we first examine the impact of road development on transport costs. Since there were very few households that reported transportation costs for production, we therefore focus on household transport costs that include costs incurred while going to such places as the market center, school, and nearest health facility. Next, we consider input and output prices such as fertilizer (urea) price, daily agricultural wage for men, as well as agricultural output and price indices.10 For the latter indices, we use the Laspeyres quantity and price indices for agricultural production. In constructing the Laspeyres indices, let k stand for commodity, i for household, and t={0,1} for year before and after the project. Defining the base year (t=0) price for each commodity as Pk , and the base year quantity of each commodity produced by the 0 household as Qk , then Qi = Qk1Pk0 k 0 is the Laspeyres quantity index for household i, and Qk0Pk0 k Pi = Qk0Pk1 k is the Laspeyres price index for household i. Qk0Pk0 k In both the RDP and RRMIMP data, the agricultural commodities entering the indices were potato, wheat, as well as high-yielding variety (HYV) Boro paddy, HYV Aman, and local and HYV Aus. Since the study covers only a base and a follow-up year, the base year indices for both aggregate price and output indices are equal to one. Note that all the values are in real terms, adjusted by the consumer price index of the base year of each survey. 9For the RRMIMP study, not all villages were visited again in the follow-up survey. A total of 7 out of 10 project villages and 2 out of 4 control villages were covered in the RRMIMP survey because (1) the roads were either not completed, or (2) control villages lost their control status as a paved road had been constructed within two kilometers of the village (thereby violating the basic assumption of being a control in the study). 10Note that similar to aggregate crop output and price indices, we could calculate an index for fertilizer prices consisting of a number of fertilizers. However, there are a few observations for types of fertilizer such as potash. So the fertilizer price here refers to the price of urea. 8 Finally, we consider household and individual outcomes such as annual total per capita expenditure, male and female labor supply, and boys' and girls' schooling. Labor supply is measured as the total number of hours in the last month worked by all men and women in household, and schooling is the percentage of school-aged boys and girls (5-17 years of age) who attended school in the year of the survey. The estimates are done separately for each project. 4. Estimates of the returns to road pavement Our primary approach in this paper is to estimate equation (3) by household fixed-effects method for all outcomes of interest, namely the transport costs of consumption, average fertilizer price, men's agricultural wage, aggregate crop output and price indices, household per capita expenditure, male and female labor supply, and boys' and girls' school participation rates. For the purpose of comparison with the fixed-effects estimates, we first present double difference-in-difference treatment effects for these outcomes across the project and control villages (Table 1). The percentage change in each outcome between year 1 and year 0 was calculated separately for project and control villages; the difference across the two groups was then calculated and tested for significance using a standard t-test. Clearly, looking at Table 1, most of the estimates are not statistically significant, with the exception of significant positive impacts for household transport costs, fertilizer price, and household per capita expenditure in RDP villages. Does this suggest that rural road investment has no significant impact on household welfare? Note that these estimates are only valid if the data are based on a randomized study or experimental design. As our study is a quasi-experimental survey design, such a difference-in-difference technique yields biased estimates. The bias reflects the difference between an experiment in which both observable and unobservable attributes have the same expectation in both treatment and control villages, and a quasi- experiment in which they do not. This can be shown as follows. From equation (3), which takes the form of a difference over time, we write the difference-in- difference equation for the treated villages (T) and control villages (C) for the two-period case as: 11 [lnYij - lnYij ] = [Hij - Hij ]+ [VT - Vj ]+ T C T C C j In theory, if the data are from a pure randomized experiment, the expected values of the bracketed terms on the right-hand-side in the above expression collapse to zero, leaving only the road impact coefficient, , which is then an estimate of the road impact. However, if the data are not from an experiment, then taking the expectation does not similarly collapse the right-hand-side bracketed expressions, in which case the estimate of will be biased upward or downward depending on how the 11We implicitly assume the road status does not change in controlled villages, while the difference takes the value 1 for the treated villages. 9 expressions on the right-hand-side turn out after the differences are performed. So in the case of our quasi-experimental survey design, the differences between the time-varying observed variables such as H and V need to be controlled for in the regression; these results, shown below, are substantially different from those of the difference-in-difference technique. The estimating equations in the household fixed-effects estimation are in semi-logarithmic form for all outcomes except for the schooling variables, where schooling is measured by the percentage of school-aged children who are in school for each household. The intervening policy variable of interest is a road investment indicator variable, equal to 1 for road project villages in year 1, and 0 otherwise (i.e. for year 0 and control villages). The road project villages include both roadside and off-road villages, while the control villages include both nearby and remote villages from the road.12 Thus, the coefficient from equation (3) measures the proportionate change in the outcome of interest (such household per capita consumption) from paving a road. That is, it represents the returns to public investment in roads in terms of paving an earthen road; the control of a paved road is having an earthen road. Note therefore that we are not comparing in this paper the impact of having a road versus not having a road of any type. The effect of having a paved versus earthen road can be shown by differentiating (3) with respect to R, where R=1 for project villages and is estimated from equation (3): (1/Yijt)Yijt /Rjt = The main estimation results for the road impact are presented in Table 2. Elasticities for the road impacts are presented in Table 3 (non-price outcomes), as well as net returns for the same outcomes (controlling for prices). As the road coefficient measures the returns to road investment, we see a high return to investment in road pavement.13 First consider the transport cost savings (Table 2). Households experience 36 percent lower transportation costs in RDP villages, and 38 percent lower costs in RRMIMP villages, because of a road development project.14 The returns to road investment in terms of transport cost savings are quite substantial. Households benefit not only in terms of transportation cost savings but also in terms of gains realized through higher prices of agricultural production, lower fertilizer prices, higher agricultural wages, 12In estimations where we allow the road effect to vary by the distance of the household from the road, note that the distance of the remote village from either the paved or unpaved road is not far enough to make another village dummy called villages without any road. Nonetheless, we interacted the distance with road status in the regression to see if the returns do vary by distance. 13We estimated two models for each outcome equation. Table 2 represents the direct estimates with no interaction; in a separate estimation we also included an interaction term between the project indicator and household distance to a paved road. Results for this second round of estimations are available upon request, as are the results for the other covariates. We discuss the relevance of the project*distance interaction in Section 5. 14Transportation costs, however, may not necessarily go down as a result of better roads; they may also increase from increased transportation frequency due to better roads. 10 and higher value of agricultural production. Table 2 also displays the estimates of road returns for input and output prices as well as agricultural productivity. Fertilizer prices are 5 percent lower in RDP and RRMIMP project villages, and real agricultural wages for men in RDP project villages are 27 percent higher than in control villages. The aggregate price index of crops is 5 percent higher and the aggregate crop output index is 39 percent higher in RDP project villages because of rural road development; similar returns accrue in RRMIMP project villages where the aggregate crop price and crop output indices are 4 and 30 percent higher, respectively, due to road investment.15 Interestingly, when computing the elasticity of the agricultural output index with respect to the road project (Table 3), we find that the elasticity is 0.27 for RDP households and 0.23 for RRMIMP households.16 These figures are very close to the road-aggregate output elasticity (0.21) obtained by Binswanger, Khandker and Rosenzweig (1993), where they use district-level data in India to examine the linkages between infrastructure, financial institutions, and aggregate output in India.17 Road development thus not only helps farmers due to higher crop production and prices, and lower prices of inputs such as fertilizer, but also benefits households via higher demand for labor, thereby raising the real agricultural wage rate of male labor (at least in RDP villages) up to 27 percent.18 Higher wages and higher demand for labor can raise family labor supply. However, the net demand for family labor also depends on the negative effect of higher income on labor supply (thus, causing higher demand for leisure), which may reduce family labor supply. Nevertheless, as Table 2 shows, male labor supply increases by 49 percent and female labor by 51 percent in RDP villages because of a road development project. There are no similar statistically significant gains in RRMIMP villages for family labor supply.19 The results do suggest that gains from both input and output markets due to road investment are substantial for rural households. The overall economic returns to road development can be measured by summing over the gains through transportation cost savings, higher output and lower input market prices, and higher productivity. While there is no an easy way we can summarize these benefits in one return estimate, such gains 15Jacoby (2000) predicts that land prices can also rise from better transport, using household distance from the market center as an indicator of transport quality in a household survey from Nepal. More specifically, he finds that land values rise by about 2% if household distance to the nearest market center falls by 10%. 16In RDP, about 68% of the villages were project villages, and about 75% of villages in the RRMIMP sample were targeted by the project. 17Normally, the effects are higher at the household level, and thus our findings are consistent with previous results for road development in comparable areas. 18This is consistent with results from Jacoby (2000) as well, who finds that wages are also lower in remote rural areas of Nepal. We also found rural electrification to have a significant negative impact on fertilizer price and increase agricultural real wage, aggregate crop price, and productivity. Interestingly, the returns to rural electrification are higher than those of road investment for some of the outcomes considered. The extended fixed effects results are available upon request. 19Lokshin and Yemtsov (2005) also find in their study of Georgia that road projects do not seem to significantly impact overall male and female labor supply in treatment as compared to control areas. 11 ultimately translate into higher household expenditure (both food and non-food), as well as human capital investment (in children, for example). The results show that the returns to road investment for household per capita expenditure are about 11 percent in the RDP and RRMIMP villages, a substantial gain in terms of higher consumption and income for rural households. This means that rural households in villages targeted by the road development project have on average an 11 percent higher consumption per capita per year.20 In addition to household consumption, average school participation among boys is about 20 percent higher in RDP villages, not controlling for proximity to the project road; road development also has a positive but barely significant impact on girls' schooling in the same sample. In the RRMIMP data, the project effect is fairly similar for boys and girls, leading to a 14 percent increase in schooling. Indeed, in our results only the project indicator seems to have had any significant impact on schooling in the RRMIMP sample. Higher school attendance of boys and girls are therefore some of social returns due to road investment; additional social returns to be assessed from road investment include returns to health and nutrition, which are not pursued in this paper. Finally, as a robustness check, we estimate the non-price outcomes (HH per capita expenditure, Laspeyres quantity index, boys' and girls' schooling, as well as male and female labor supply) on the same set of explanatory variables as well as the price variables discussed above (men's agricultural wage, fertilizer price, and the agricultural price index facing the household). The net returns are presented in table 3, and reflect the returns if we held the structural model (taking prices as exogenous) to be correct. Overall, the estimated effects are slightly higher than our original returns, except for the aggregate output index where the return is lower. The results are fairly robust and in almost all cases retain significance with our earlier estimates.21 5. Distribution of benefits of road investment While the above analysis allows for a broad range of controls in road project evaluation to generate average returns to road investment, it still imposes a single response coefficient on all households and communities conditional on other factors, such as the household distance to a paved road. The estimates of the earlier section are based on the assumption that distance does not matter in terms of gains from a road development project. This may not be the case, however. That is, households located in a roadside village may benefit more from road investment than households living away from a paved road. Testing this hypothesis requires differences in household distance from road, which is very minimal 20The returns to consumption seem to be higher for rural electrification than rural road investment (as high as 23 percent for electrification as compared to 11 percent for rural road development). 21The only exceptions are boys' schooling and female labor in the RDP sample. 12 by the fact that remote villages in both project and control road areas are selected on an almost equal distance basis, at least more so in RDP than RRMIMP areas. Yet when we have interacted road distance and its square with the road status variable, we find that not surprisingly, except for boys' schooling, the distance to a developed road does not matter at all in both RDP and RRMIMP villages for any outcome considered. In the case of boys' schooling, the effect of road development diminishes with the distance at an increasing rate.22 Gains from a road development project may also vary by household income status. In this paper, we examine the distributional issue of gains from road investment using quantile regression analysis. It is potentially important to investigate changes in outcomes observed at different points in the income or consumption distribution. Simply investigating changes in the mean may not be sufficient when the entire shape of the distribution changes significantly (Buchinsky, 1998). Studying the distributional impact also sheds light on political constraints on the allocation of infrastructural investment (Jacoby, 2000). Following the model proposed by Koenker and Bassett (1978), assume yi , i = 1,...n, is a sample of observations on the log consumption, and that xi is a K x 1 vector (comprising the project (R), household (H), and village (V) level characteristics controlled on the right-hand side of equation (3)). The quantile regression model can be expressed as: yi = xi + i, Q (yi xi ) = xi , (0,1) ' ' (4) where Q (yi xi) denotes the quantile of log per capita expenditure conditional on the vector of covariates (x). In general, the -th sample quantile of y solves: min 1 (5) yi - xi +' ' n i:yi xi ' i:yi