Rural Roads and Local Market Development in Vietnam

Abstract We assess impacts of rural road rehabilitation on market development at the commune level in rural Vietnam and examine the geographic, community, and household covariates of impact. Double difference and matching methods are used to address sources of selection bias in identifying impacts. The results point to significant average impacts on the development of local markets. There is also evidence of considerable impact heterogeneity, with a tendency for poorer communes to have higher impacts due to lower levels of initial market development. Yet, some poor areas are also saddled with other attributes that reduce those impacts.


I. Introduction
The literature on rural roads and economic development has emphasised impacts on transport costs and prices with consequent welfare impacts. For example, rural roads may allow farmers in remote (and often poor) rural areas to obtain higher prices for their outputs, and/or reduce the prices of inputs and consumer goods. 1 However, this says little about how rural roads might also influence the geography of economic activity and, in particular, what role road improvements might play in local market and market-related institutional development. Initial conditions in remote poor areas are often characterised by highly geographically incomplete and non-existent markets. The goods concerned are simply not available in these areas, given high transport costs. Advocates of rural road projects often point to their potential benefits in stimulating market development ). Yet, rigorous evidence appears to be non-existent.
It is important to distinguish two ways in which access to markets can improve due to better transport infrastructure. One is through reduced travel costs to existing markets and institutions. The other is through the induced relocation of markets (as physical entities) and institutions. The bulk of the literature on rural roads appears to have the first in mind. However, markets are mobile -not least in developing countries. One response to road improvements could be the development of local markets defined as fixed places where villagers and outsiders gather at established times to buy and sell goods.
Why should we care about whether the residents of a poor area have goods commercially transported to their community and available in a local market, rather than travel themselves to an outside market? At a purely descriptive level, it may be of interest to know how road improvements affect the geography of economic activity. Does economic activity in a developing country become more geographically concentrated or less so as roads improve? There may also be important instrumental reasons. While it should not be presumed that local market development is welfare enhancing, supportive arguments for that view can be made. One possible reason is that there could be large external benefits to having a local market.
There are two relevant literatures here. The economic geography literature has postulated that externalities -agglomeration economies -play a crucial role in the spatial concentration of economic activity. 2 A local market's physical presence and facilitation of trade could be an instigating factor in a process of shifting production structures to more diversified and higher value activities, improved access to various services, and broader economic development in an area. 3 Such benefits are external in the sense that decisions by the commercial carrier to transport goods to the community or not will not take them into account. Local market development may then create a virtuous circle whereby the stimulation of off-farm development and new income earning opportunities result in higher perceived returns to education and in time, higher schooling. We will dub this the hypothesis of 'transport-induced local-market development (TILD). ' A second set of reasons for believing that local market development brings local benefits is found in the (largely non-economic) literature on the role of local markets as forums for the exchange of ideas and learning, recreation and social interactions (see Skinner, 1965Skinner, , 1985Masschaele, 2002;Liu, 2007). This literature emphasises the role of local markets in social change and in connecting isolated rural peasant communities to the external social and economic system. Others have written about the importance of trust and relationships in exchange as might be expected to foment in local rural markets (for example, Fafchamps and Minten, 1999). These are about production consequences as well as non-economic benefits of social development.
However, it is far from clear that public investments in transport will actually promote local market development in poor areas. In its analysis of where economic activities take place and why, the new economic geography points to increasing returns to scale leading to agglomeration economies. This may make it hard for markets to develop in poor areas even with large reductions in transport costs. Theoretical arguments can also be made that under certain conditions, road improvements could either encourage local market development or discourage it as local residents in the targeted areas can now more easily reach established markets .
This paper aims to throw light on these issues and to test TILD, by assessing the impacts of a World Bank-financed rural road rehabilitation project implemented in Vietnam between 1997Vietnam between -2001 Numerous observers of the rural Vietnam setting have remarked on the correlations between road infrastructure, local markets, and offfarm income diversification (Kerkvliet and Porter, 1995;Bryceson et al., 2008;Minot et al., 2006). And indeed, the objectives of the project were to develop local market activity and, hence, economic development through targeting road improvements to poor communes.
We focus on whether the road improvements lead to the development of local markets and whether there are also impacts on off-farm development consistent with TILD. Our data and methods allow us to assess this in a methodologically rigorous way -controlling for how road sub-projects are allocated and for initial conditions that may affect subsequent outcome trends, and monitoring communes long enough to capture impacts.
We concentrate on three sets of questions. The first concerns average impacts on local market development. The project coincided with a period of rapid economic development with access to markets of various kinds increasing substantially over the study period. How much of the change observed in project communes can be attributed to the road project? A related issue is how impacts vary over time. If the transport cost saving is large enough and the agglomeration effects noted above are present, then we would expect project impacts to increase over time. Evidence on this point is scarce. 5 Our second set of questions concern cross-commune differences in the project's impacts on local markets and what explains those differences. 6 Heterogeneity of impacts can be expected to arise according to the economic and social characteristics of the communities. In this context, an important issue for project design is whether higher levels of initial development enhance or diminish impacts. Should governments and donors target places that have the market institutions necessary for further economic development or focus resources on the places without such attributes? In practice, project selection often tries to favour poor areas with poor road conditions. The fact that poor areas tend to have less market development to begin with suggests greater potential for roads to enhance market development in poor areas. However, poor areas are typically also saddled with attributes that may well prevent markets from developing. Depending on precisely how road benefits depend on initial conditions, project design may need to consider complementary inputs and policies to achieve the potential benefits.
This leads to our third question, which concerns the structure of the crosscommune differences in impacts. It is well recognised that the same intervention can have different impacts on different places and households. However, the policy implications of such heterogeneity depend crucially on whether it shares a common structure across outcome variables. If communes with better educated households derive larger impacts for improved access across all kinds of markets, for example, then a robust conclusion can be drawn about the gains from targeting such commune covariates. If the relevant sources of cross-commune differences in impacts vary greatly across multiple outcomes then it will be hard to exploit heterogeneity to assure better projects.
Section II briefly describes the project being evaluated, our data and initial conditions in our sample of rural communes. Section III discusses our methods and Section IV our results, while Section V concludes.

Rural Roads and Local Market Development 711
II. The Project, Data and Setting

The Rural Road Project Intervention
The Vietnam Rural Transport Project I (RTPI) aimed to link commune centres to markets, stimulate market development and reduce poverty through the rehabilitation of 5,000 kilometres of rural roads (World Bank, 1996). The project was implemented between 1997-2001 in communes located in 18 provinces scattered around Vietnam. Participating provinces were responsible for choosing communes for inclusion in RTP1, as well as the road links to be rehabilitated within them. On paper, road links were identified through least cost techniques, and eligible if the road's zone of influence had a population of over 300 per kilometre, and average rehabilitation costs were below $15,000 per kilometre. 7 In mountainous communes with a high density of ethnic minority households, provisions were made for the possible waiving of the population and cost criteria. In practice these eligibility criteria identify considerably more road links than could be covered by the project. How the included links were selected among these is unclear.
Many of the targeted roads were in very bad condition, some with impassable sections year round. A rehabilitation standard of 'reliable access' was enforced that provides relatively consistent and safe access with only short-term road closures (due to bad weather). The project expressly stipulated that no 'new' roads would be built.
Aid or central government spending for road projects may substitute for local government spending intended for the same purpose, by being diverted to other sectors or to neighbouring non-project areas. Elsewhere we have ascertained that the project did produce differential impacts on the kilometres of improved roads in project compared to non-project comparison communes (van de Walle and . We found no evidence that resources were diverted to non-project communes for roads or other basic infrastructure. However, we also showed that project funds were used not only to rehabilitate roads as intended by the project, but also to build new roads. The impacts we study in this paper are due to both types of improvements.
Using the methods to be described in section III, we have also checked to see whether the period under analysis was marked by differences in the implementation of other development projects in the communes. For a long list of potential interventions for which information is available in our data, we find no evidence that project or non-project communes were treated differently. 8 Based on the findings, we are confident in attributing any differences in outcome changes over time to rural road rehabilitation and construction.

The SIRRV Data
Collected specifically for evaluating the impacts of the rural roads rehabilitated under RTPI, the 'Survey of Impacts of Rural Roads in Vietnam' (SIRRV) consists of a panel of 200 communes and 3000 households. The survey design implicitly takes the commune as the project's zone of influence. This is justified by the project objectives -namely to link commune centres (where key social, economic and administrative facilities are located) with road and market networks -and because the commune is an enumeration level at which data is commonly collected in Vietnam.
The baseline was collected pre-project starting in June 1997, while subsequent rounds followed in the summers of 1999, 2001 and 2003, tracing the implementation process and schedules of prior rounds. The analysis makes use primarily of the first and last rounds, though we will also test impact dynamics using the 2001 round.
Project ('treatment') and non-project ('comparison') communes, and households within them, were surveyed in six of the 18 provinces participating in the project -Lao Cai and Thai Nguyen in the north, Nghe An and Binh Thuan in the centre, and Kon Tum and Tra Vinh in the south of the country. 9 Project communes were randomly selected for inclusion in the survey from province-specific lists of all communes with road links proposed for rehabilitation. Another list was drawn up of remaining communes in districts with proposed road links from which a random sample of non-project communes was drawn. The eventually sampled communes (100 project and non-project each) were located in 29 of 38 potential survey districts.
Non-project communes located in the same districts as the treatment communes will share many of the same characteristics as the project communes. Yet we additionally use matching techniques to ensure selection of the most appropriate comparison communes. Districts are large and the distances between project and non-project communes also tend to be large. 10 Contamination from one to the other is unlikely for the type of small localised road improvements under study. The analysis focuses on commune level impacts and primarily uses data collected at the commune level. In each sampled commune, 15 households selected through a stratified sampling framework also answered a household questionnaire. 11 It does not measure income or consumption. However, using extensive information on household characteristics common to the SIRRV and the nationally representative Vietnam Living Standards Survey (VLSS) of 1998, we use regression techniques to predict consumption expenditures for SIRRV households in 1997. 12 This variable is then aggregated to form a commune level welfare indicator which we use to divide communes into those below (poorer) and above (better-off) median mean predicted consumption.
Finally, a project level database detailing what the project did, when and how, was also constructed for each surveyed project area. Project differences can then be taken into account in determining impacts.

Initial Conditions of Poorer and Better-Off Communes
The pre-project setting can be described as one of relatively isolated, primarily subsistence-oriented villages with generally poor market related infrastructure -a legacy of the communist period that suppressed markets, and of their slow development since the State and the farming collectives relinquished control of the economy in the late 1980s (Fforde and de Vylder, 1996). Continuing restrictions on mobility, the lack of safety nets and the threat of losing land that is left uncultivated, render households in these areas relatively immobile. In similar areas of China, local markets have been described as providing a meeting place to engage in trade, business and social interactions for small-scale cultivating households scattered over large areas (Skinner, 1985;Liu, 2007). Itinerant traders and assemblers can buy small Rural Roads and Local Market Development 713 quantities from peasants. At the same time the market allows peasants to obtain needed goods and services (Luu, 2003). Yet, unlike similar rural areas in China where village and township enterprises have played a vital role in off-farm diversification, rural Vietnam has had no such tradition. Thus the role of local markets may be commensurately more important in Vietnam.
Pre-project, 48 per cent of all sample communes had a market whose frequency averaged once a week. Sixty seven per cent were open air with no building; a quarter of the market buildings were of brick, while the rest were described as 'temporary.' The market areas ranged from 72-2500 square metres. The vast majority had been built recently, since the collapse of the collective farming system. Figure 1 shows how the initial presence of a commune market was related to both the commune's distance to the closest central market town and its average living standards as measured by mean predicted household per capita consumption in the baseline. 13,14 Panel (a) plots the relationship between having a market (vertical axis) and commune mean consumption on the horizontal axis for communes classified into three equal groups based on distance to the closest large market town -within 7km ('close'), between 7-15km ('middle') and further than 15km ('far'). Panel (b) places distance on the horizontal axis instead and plots the relationship separately for communes above and below median consumption. Figure 1 suggests a clear relationship between these three variables. Local markets are most often present at middle distances. They are an increasing function of predicted consumption except at very close distances and high consumption levels (Panel a). For better-off communes the relationship with distance is a pronounced inverted U, while it is much flatter for poorer communes although starting out at a somewhat higher level for those at very close distances (Panel b). Poorer communes are less likely to have markets than better-off communes at all distances beyond short distances (Panels a and b). Table 1 examines how communes below and above median consumption expenditures differ in their initial characteristics -including (in the bottom half) the market and market-related development outcome indicators that we focus on. 15 The table reveals considerable and highly significant differences in attributes across communes disaggregated in this way. Poorer communes are associated with characteristics that are typically assumed to be disadvantageous, including higher illiteracy, worse access to transportation and credit, larger distances to the closest city and far lower market presence. They generally have lower levels of population and road densities, larger minority populations, and are more likely to be in mountainous areas.
Focusing specifically on the baseline values of the outcomes variables, we see that in addition to having lower market presence, poorer communes typically have significantly fewer commercial businesses, inferior access to services, less diversified income sources and worse schooling indicators. For example, only 32 per cent had any kind of market, and small shops or stalls which typically sell a few basic necessities such as salt, matches and soap, were present in only 39 per cent. The probabilities that better off communes had markets and shops were 63 and 57 per cent, respectively. In 1997, an overwhelming majority of households in these communes relied primarily on agriculture for their livelihoods (90%) but the lack of income diversification was even more pronounced in the poorer communes where 94 Notes: These are non-parametric regressions, using locally weighted smoothed scatter plots. The unit of observation is the commune. One third of all communes are 'close' or within 7km of a central market; the 'middle' distance third are between 7-15km from the market; while the 'far' communes are more than 15km from the market. Better-off and poorer communes are defined as those above or below the sample median commune per capita consumption based on aggregated predicted household consumption.
per cent did so, compared to 86 per cent. Finally, less than a quarter of children completed primary school by age 15 years in the poor communes and only threequarters of these continued on to secondary school. In the non-poor communes it was 36 and 92 per cent, respectively. These differences again raise the crucial policy question of whether road placement in poor areas with poor initial conditions will handicap or stimulate TILD. The rest Notes: The sample consists of all 200 communes. **significant at 5 per cent level or higher; *significant at 10 per cent level. Flood and storm prevalence summarises the average incidence between 1997-2003; credit availability averages dummy variables for the availability of credit from the following sources: the Agricultural Bank, commercial banks, the Bank for the Poor, credit co-ops/ people's credit funds, government programmes, mass organisations, international projects and NGOs; transportation accessibility averages dummy variables for the presence of provincial and national roads, railways and waterways. Many outcome variables are dichotomous referring to whether the outcome is present in the commune. The exceptions are: market frequency which takes the values 0 for no market, 1 for once per week or less, 2 for more than once a week, and 3 for permanent market; the percentage of households in various occupations refers to their main source of income; the primary completion rate is defined as the share of children aged 15 years and under who completed primary school; the secondary school enrolment rate is the share of children who graduated from primary school in the previous year who are enrolled in secondary school.
of the paper explores whether the project had impacts on these outcome indicators and how differences in initial conditions may have interacted with road improvements to affect those impacts.

III. Evaluation Methodology
The official project selection criteria detailed in Section II clearly allow provinces considerable freedom in choosing communes and road links. Consequently, the placement of the project is unlikely to have been random and may well have been influenced by factors that also determine outcomes.
A potentially important source of endogeneity bias in this context is that initial conditions are likely to determine project placement, as well as to influence the subsequent growth path and prospects of the communes (Jalan and Ravallion, 1998). Our evaluation methodology aims to correct for these potential sources of selection bias.
We combine a difference-in-difference (DD) with propensity score methods (PSM). A conventional DD gives unbiased estimates based on the assumption that the selection bias is constant over time. However, if there are time varying factors that influence placement, then road placement is still correlated with the error term in the differenced equation. To allow for the possibility of time variant selection bias due to initial observables, we use the predicted probability of participating in the road project (the propensity score) to match the comparison communes in the DD estimate. The propensity score is estimated using a logit that includes initial conditions that may affect subsequent commune trajectories as explanatory variables. Our impact estimates are then constructed by comparing the before and after project change in outcome measures for the project communes with those for the matched comparison communes.
Specifically, the average impact for project communes (DD) can be written as: where is the impact estimate for commune i, P and NP denote project (treatment) and nonproject (comparison) communes respectively; Y P i1 À Y P i0 is the change in the outcome measure for project commune i; Y NP j1 À Y NP j0 is the change in the outcome measure for comparison commune j; and W ij is the weight given to the jth commune in making a comparison with the ith project commune. N p in Equation (1) is the total number of project communes. We apply nonparametric kernel matching in which all the nonparticipants are used as comparison communes and weights are assigned according to a kernel function of the predicted propensity score following Heckman et al. (1997). This technique ensures valid bootstrapped standard errors (Abadie and Rural Roads and Local Market Development 717 Imbens, 2006). 16 As a robustness check, we also construct a PS-weighted DD (Hirano and Imbens, 2002;Hirano et al., 2003). 17 The key assumption of PS-matched or weighted DD in this context is that the selection bias is conditional on the observed placement covariates in the baseline. The estimates will be biased if there are unobservables that affect both project placement and outcome changes. All project communes were selected prior to the project start date based on initial conditions as reflected in our baseline. The logit model used to calculate the propensity scores controls for an array of initial conditions that may subsequently affect changes in the communes. However, we can never rule out the possibility of omitted initial conditions that are correlated with placement and outcome changes over time.
To explore whether and how initial commune conditions affect impacts we use a simple ordinary least squares (OLS) regression of the estimated commune specific impacts against commune characteristics. For this exercise we use the PS matched estimates since these can be estimated for each specific commune.

IV. Impacts on Local Market Development and their Heterogeneity
Before examining average impacts and impact heterogeneity, we briefly discuss the estimation of the propensity scores used to match project and non-project communes.

Participation in the Project
The probability of a commune's participation in the project is estimated using a logit model. 18 We find a number of significant explanatory variables for programme placement. Consistent with official selection criteria, communes with a higher total population and a larger share of ethnic minority population were more likely to participate in the project. A few characteristics that may indicate higher living standards had a significant negative effect on the probability of participationnamely, the share of the adult population working in private enterprises, the school enrolment rate and having an Agricultural Bank branch. Yet, other proxies for income had no effect -including, the presence of a market and predicted average commune consumption expenditures. Finally, among measures of transport and accessibility, a national road passing through the commune, the presence of passenger transport, a higher density of roads and distance to the province centre all reduced the probability of participation. As there is imperfect overlap in the estimated propensity scores, we limit the sample to the common support, ending up with 94 project and 95 non-project communes for the rest of the analysis. Matching on the predicted propensity scores, we achieve a close balancing of the initial observed commune characteristics for the two samples. 19

Average Treatment Effects
We assess impacts of the road project on a set of outcome variables (introduced in Table 1) that we deem relevant to local market development. In addition to the presence and frequency of local markets, we examine whether the presence of other commercial establishments -namely shops, bike repair shops, pharmacies, and 718 R. Mu & D. van de Walle restaurants -was affected. Such impacts could both be direct or via impacts on local markets. To test whether there are signs of a process consistent with lower transport costs and market development stimulating a more diversified local economy, we also examine whether there are effects on the availability of various services and signs of livelihood diversification away from agriculture and towards trade and service activities. Finally, we check to see whether school enrolments are affected, as might be the case if the perceived returns to education have been altered. Table 2 displays the mean values of these indicators across project and non-project communes in the baseline and for subsequent survey rounds. 20 These generally moved in the expected direction over time, with a tendency to increase in both commune groups. The key question then is whether there was a differential impact attributable to the road improvements in the project communes. Table 3 presents DD estimates of the mean impacts using the PS-based kernel matching and weighting methods discussed in section III, as well as simple DD estimates. The estimates are given for 1997-2001 and 1997-2003, referred to as the short and medium term. Under our assumptions, these estimates reflect causal effects Notes: The sample consists of the 94 project and 95 non-project communes on common support as determined by propensity score matching. Many outcome variables are dichotomous referring to whether the outcome is present in the commune. The exceptions are: market frequency which takes the values 0 for no market, 1 for once per week or less, 2 for more than once a week, and 3 for permanent market; the percentage of households in various occupations refers to their main source of income; the primary completion rate is defined as the share of children aged 15 years and under who completed primary school; the secondary school enrolment rate is the share of children who graduated from primary school in the previous year who are enrolled in secondary school.

Notes:
The sample consists of the 94 project and 95 non-project communes on common support as determined by propensity score matching. T-ratio of kernel matching is obtained from bootstrapping (100 repetitions). ** significant at 5 per cent level or higher; * significant at 10 per cent level. Standard errors of weighted DD estimations are robust to heteroskedasticity and serial correlation of communes within the same district.
of the road improvements. One or two stars indicate whether each change is significantly different from zero at the 10 and 5 per cent significance levels, respectively. By the start of data collection for the 2001 round, 27 months had elapsed on average since the project work ended. 21 How long it takes for impacts to emerge is an issue that often arises in discussions of road impacts and planning for their evaluation. Here we are able to ascertain whether local area impacts were different in 2001 from those in 2003, after two more years had elapsed.
Focusing on the PS-based estimates, and starting with impacts by 2001, we see that across the examined indicators there is no sign of statistically significant mean impacts in the short-term. The only exception is for the primary school completion rate which rose by 15 to 25 per cent, according to the kernel matched and PSweighted DD respectively. Why would better roads affect primary school completion rates? The road may make the school more accessible to children in the commune's outlying villages. And, although all communes have primary schools, secondary schools are considerably rarer. It is plausible that a road improvement allows children to more readily reach a secondary school which will encourage both primary school completion and post primary enrolments.
The results change when we track impacts through to 2003. A number of outcome indicators now exhibit significant impacts. As a result of the road improvements, markets became newly available in close to 10 per cent more project than non-project communes over the seven years, and their frequency increased. However, despite small positive impacts on commercial establishments, none are statistically significant.
By 2003, we also discover significant impacts of the road project on the services for which we have data -the availability of tailoring and hairdressing services. The weighted DD show that the probability of men and women's hairdressing services being available in the communes rose by 14 and 20 per cent respectively, in 2003. Consistent with effects on market and services availability, we find evidence of impacts on employment and livelihood patterns. Improved roads resulted in a small but significant 2 per cent decline in households relying on farming as their main source of income. A significant increase in the share of households mainly relying on the service sector (1.7 per cent) hints at what alternative livelihoods these households may have switched to. This is not a trivial impact given that only 1 per cent of households were employed in the service sector in the baseline. The impact on households engaged primarily in trading activities is also positive but small and statistically insignificant. Finally, impacts on the primary school completion rate are sustained over time and have even risen slightly. Moreover, small effects on secondary school enrolments also appear to be emerging.
In sum, we find some support for TILD. Our results indicate significant average impacts on the development of local markets, for both their presence and frequency. The project resulted in households switching from agriculture to non-agricultural, mostly service-related activities. Tailoring and hairdressing services became more commonly available. The impacts were not sharp and short-lived; they took time to emerge, only appearing in 2003, and are thus rising over time. This is all we can say based on our two data points. Of note too are the quicker, sustained and robust impacts on primary school completion rates.

Heterogeneity in Impacts
As implied by our theoretical model, the average treatment effects may hide significant heterogeneity across communes. Using the PS-matched DD method and focusing on the 2003 data, we calculate the individual treatment effects for each of the 94 project communes. Eyeballing these confirms that they vary substantially across communes. Furthermore, calculating mean impacts separately for the 47 communes below and above median predicted household consumption reveals pronounced differences in impact estimates between relatively poorer and less-poor project communes. Particularly striking is that impacts are generally larger for the poorer communes. Normalising impacts by each group's mean value of the variable in the baseline, we find that for 10 out of the 14 outcomes the impacts for the relatively poorer communes exceed those for the better off ones. 22 The characteristics associated with whether a commune is poor or not are likely to interact with roads to influence their impacts. One popular hypothesis is that benefits are highly dependent on local human capital endowments needed to take advantage of the opportunities afforded by new roads. However, our finding that impacts are larger in poorer communes where, as we saw in Table 1, illiteracy is also typically higher, appears to contradict this common argument. Other hypotheses can be suggested, such as historical discrimination against certain social and economic groups makes it harder for them to adopt more outward economic orientation, as required to take advantage of new roads. Our result of generally higher impacts in consumption poor communes -which also tend to have worse attributes -begs for analysis of the role of initial conditions in determining road impacts.
To explore the covariates of road impact estimates, we use OLS regressions where the dependent variables are the commune level impact estimates and the explanatory variables are initial pre-project commune characteristics for the sample of 94 treatment communes. Potentially important, mediating physical, social and economic commune conditions, that we also observe, include most of the variables listed in the top half of Table 1. To these we add location in the country's northwhich has had a far shorter experience with the market economy. We include the initial value of the dependent variable/outcome measure, as well as whether the commune had a local market pre-project, as a test of the virtuous cycle idea. Finally, to represent heterogeneities in the actual treatment we also include quadratics in the number of months since project completion and in the length of improved road. 23 In principle one can imagine all sorts of relationships and non-linearities between these project attributes and impacts. Time may enhance impacts as local providers take time to set up or it may reduce them as customers come to value access to outside providers. Under increasing returns to scale, one would expect cumulative impacts with more time leading to higher impacts.
The interpretation of road length is a bit unclear, though it is still probably better to control for it as it represents an important difference in treatment. Typically, the project rehabilitated what was necessary to make the road link functional. Length is thus likely to reflect some omitted characteristic about how bad road access was prior to the project. It is probably not interpretable as road length per se but most likely proxies for the road's initial condition and omitted attributes of remoteness. Tables 4-8 report the results. 24 For each outcome variable we present two regressions: one with a full set of the same initial conditions (model 1) and one the result of a cumulative pruning of the highly insignificant variables (t-statistics below one), starting with the lowest t-ratio (model 2). This serves to sharpen the picture somewhat, given multicollinearity.
There are significant interaction effects, indicating that impacts are the result of how the attributes of places and people interact with what the project does. Some attributes consistently raise or reduce impacts, while a few are both complements and substitutes to better roads in inducing local market development. . T-statistics are given in parentheses. **significant at 5 per cent level or higher; *significant at 10 per cent level. Market is a zero/one dummy for whether a market exists in the commune. Market frequency takes the value 0 for no market; 1 for once a week or less; 2 for more than once a week and 3 for permanent market.

Notes:
The dependent variables are the 94 estimated commune specific impacts for 2003. Standard errors are clustered at the district level of which there are 29. T-statistics are given in parentheses. **significant at 5 per cent level or higher; *significant at 10 per cent level. All outcomes refer to availability in the commune.

Rural Roads and Local Market Development 725
We note first that impacts are consistently and significantly reduced for communes with a higher initial value of the outcome variable. 25 These are some of the largest effects both in terms of the magnitude of the coefficients and of their statistical power. This strongly suggests decreasing impacts -whereby marginal returns are higher when outcomes are initially lower. This is consistent with our earlier finding that impacts tend to be higher in poorer communes.
As anticipated, several commune attributes that are widely deemed to be disadvantageous, consistently dampen the impacts of improved roads, although . T-statistics are given in parentheses. **significant at 5 per cent level or higher; *significant at 10 per cent level. All outcomes refer to availability in the commune.
not significantly across all outcomes. For example, as we would expect, higher adult illiteracy rates reduce the impacts of road improvements on a number of market related outcomes -the presence of commercial establishments, the availability of services and secondary school enrolments -consistent with human capital and infrastructure being complements. Illiteracy also strengthens road impacts on the share of households who remain farmers. A greater distance to the closest market town significantly lowers impact on the availability of pharmacies, tailoring and women's hairdressing services, specialisation of households into the service sector and secondary school enrolments. The last probably reflects the fact that distance to secondary schools is closely correlated with distance to the market town. As expected, impacts are also generally lower for communes located in the North where entrepreneurship and markets have been less developed historically.
A high concentration of ethnic minority households, controlling for a mountainous location and education levels, results in significantly lower impacts on many of the same outcomes including markets and their frequency, the availability of services and continuation on to secondary school. This too may reflect the fact that many minorities have less of a tradition of using markets or relying on public services due to a culture moulded by past discrimination. This is broadly consistent with the arguments and evidence of van de Walle and Gunewardena (2001) on the sources of ethnic inequality in Vietnam.
Other commune characteristics have almost exclusively positive effects on the impacts of improved roads. The initial presence of a market in the commune typically significantly enhances road impacts on other market-related development consistent with a story of external benefits to local markets and the hypothesis of TILD. Unsurprisingly, initial market presence particularly enhances impacts on the establishment of retail and other small firms, and trading activities. It also significantly increases impacts on primary school completion rates.
Population density is typically a project placement criterion as indeed it was for RTP1. Impacts and marginal returns are expected to be higher in more densely populated communes. We find supportive evidence for this with respect to impacts on shops and women's hairdressing services. More households with motorcycles indicate the degree to which households can rapidly take advantage of the road for their transport needs, although it may also capture an income effect. Plausibly, we find that it enhances project impacts on the development of off-farm activities and secondary school enrolments.
A number of other commune attributes interact with road improvements to both raise and reduce impacts depending on the outcome indicator. A much cited bottleneck to development is lack of credit. We find evidence for this with respect to the development of local markets. However, credit availability appears to reduce road impacts on household diversification into trade and service sector activities. Credit has been found to be more readily available to landed households engaging in agricultural pursuits in rural Vietnam (Ravallion and van de Walle, 2008: chap. 7). Its availability may well coincide with other discouragements to trade and service sector activities.
A high prevalence of weather shocks and presumably a higher incidence of episodes of commune inaccessibility significantly reduce the impacts of road Rural Roads and Local Market Development 727 improvements on the availability of shops and school enrolments at the secondary level. Against this, it raises impacts on primary completion rates and on the share of households relying on the service sector for their livelihoods. Location in mountainous areas also reveals ambiguities in its impacts across outcomes. It significantly increases road impacts on local market development and the availability of tailoring services, reinforcing our intuition that, holding other attributes constant, poor road conditions represent a key constraint to market development in mountainous areas. Yet, mountainous location also interacts with the project to reduce the percentage of households who derive their livelihoods from farming and services, as well as the primary completion rate. The number of months since project completion has both positive and negative impacts on a number of service-related indicators. Restaurants are more likely to develop as more time elapses. More months also have a positive though decreasing . T-statistics are given in parentheses. **significant at 5 per cent level; *significant at 10 per cent level. The primary completion rate is defined as the share of children 15 years and under who completed primary school; the secondary school enrolment rate is the share of children who graduated from primary school in the previous year who are enrolled in secondary school.

Rural Roads and Local Market Development 729
impact on the share of households relying on the service sector. On the other hand, the longer the period, the lower the impact on the availability of tailoring services. Road impacts on women's hairdressing are also first negatively affected by more time passing but this is reversed after around 50 months have gone by. Finally, the length of improved road is significant in a number of cases but, as anticipated earlier, its interpretation is unclear.

V. Conclusions
We study the impacts of rural road improvements on local markets and marketrelated development at the commune level in Vietnam. In particular, we endeavour to test whether impacts are consistent with a hypothesis of 'transport-induced local market development (TILD).' Our empirical methods combine a double difference estimator with propensity score matching on pre-intervention covariates. We examine average impacts, including the time it takes for those impacts to emerge and whether they rise or fall over time but also the cross-commune differences in impacts, and the nature of those differences, including interactions with initial geographic, community and household characteristics. We focus on two specific questions that are important from a policy viewpoint, to see what implications there might be for future project design. Are road impacts enhanced or weakened by initially poor local market development as is typical in poor areas? Are the covariates of road impacts congruent across outcomes? These issues have tended to be ignored by the literature on assessing rural road impacts. There are indications of significant average impacts on the development of local markets and related indicators. Few outcomes responded rapidly to the new and improved roads. Most impacts are not apparent 27 months (on average) after project completion, and only emerge in data collected two years later. We find significant average impacts on the presence and frequency of markets and on the availability of various services. The project also resulted in households switching from agriculture to non-agricultural, mostly service-based, activities. Perhaps most notable, the project had significant, early and sustained impacts on primary school completion rates. These results give qualified support for the hypothesis of TILD.
However, it is clear from our findings that TILD oversimplifies the process. Our findings point to substantial heterogeneity in the effects on market development. The circumstances of a project's location influence its impacts. On the whole, poor communes tend to experience higher impacts on many indicators of market development. This is the outcome of two broad sets of attributes of poor areas that tend to work in opposite directions to influence the impacts on local markets of road improvements. On the one hand, poor areas are less likely to have markets and market-related institutions and services and this alone means more scope for road improvements to help develop those same institutions and services. On the other hand, poor areas have various other attributes that tend to discourage TILD. Poor communes in Vietnam, for example, are more likely to have a high share of ethnic minorities and high illiteracy rates which have negative effects for most outcomes. They are more isolated and have lower population densities, attributes that also tend to lessen road impacts. They are less likely to initially have a local market which impedes development of other market-related institutions and services in response to road improvements (separately to the fact that markets are more likely to develop in places where they do not exist initially). Hence, we find signs of a virtuous cycle whereby the impacts on small businesses, service availability, trade activities and primary school completion rates are enhanced by the initial presence of a market.
Our results suggest that, on balance, the road project tended to have larger impacts on market development in poorer communes due largely to the initially lower market development in these places. This was strong enough to outweigh the fact that poorer communes have other attributes (besides low initial market development) that reduce impacts of road improvements.
The structure of the cross-commune heterogeneity in outcomes is driven by the initial state of market development tempered by a number of commune attributes in a way that tends to follow distinct and predictable patterns across outcome indicators. Distance to central markets, low population density and high minority populations, high adult illiteracy and location in the North all consistently dampen road impacts.
These findings can be exploited by project design to promote larger development impacts. They suggest that small road improvement projects such as the project studied here could have vastly larger impacts on local market development if they were targeted to places with initially lower market development, and equally important, accompanied by complementary social and economic policies aimed at improving certain attributes (such as adult literacy) or reducing the disadvantages of others (policies to reverse the effects of historical discrimination towards ethnic minority groups) that interact with roads to reduce their impacts.