Spatial Networks, Incentives and the Dynamics of Village Economy: Evidence from Indonesia 1

Transcription

1 Spatial Networks, Incentives and the Dynamics of Village Economy: Evidence from Indonesia 1 subdistrict Keywords : Income Growth, Spatial Connectivity, Education, Indonesia Introduction Economic growth often accompanies spatial inequality. Spatial connection to high growth centers promises the pathway from poverty in local economies, improving economic returns to investment with reduced transportation and search cost for both human and physical resources, which alters the household resource allocation. In general, the improvement of spatial connectivity is expected to increase allocative efficiency in the local economy, since therefore the mobility of resources becomes faster and less costly and thus price disparity becomes smaller (e.g., Minten and Kyle, 1999). However, the implication of better spatial connectivity among neighborhood local areas and/or with distant economic centers on poverty and income distribution is not clear. In rural contexts, once a village is connected by a new road to the nearby local town where jobs are available, the labor allocation is expected to change so that households try to gain from the town labor market. If entry to the labor market is easier only among educated agents, the labor This study is based on collaboration between the Japan Bank for International Cooperation Institute (JBICI), the International Food Policy Research Institute (IFPRI), and the Indonesian Center for Agriculture and Socio Economic Policy Study (ICASEPS). The authors thank Yuki Takako and Shimokoshi Shinobu for research assistance.

2 allocation changes among households with educated members. More educated agents may try better employment or urban market opportunities available at a larger economic center farther than the local town (without migration). In this case, road access to the larger economic center is more important. Similarly, if the local town has efficient agricultural product markets, landed farmers will benefit from the new local road, not landless households, because farmers have product surplus to sell in the markets. Increasing demand for food from larger economic centers may induce some landed farmers to invest in agroprocessing, increasing nonfarm income. Therefore, the above example implies that the effects could be heterogeneous across different locations and across households with different endowments. The recent literature provides some studies suggesting that returns to human and physical capital in rural areas critically depend on spatial connectivity, which affects the household resource allocation such as labor supply (e.g., Fafchamps and Shilpi, 2003, 2005; Fafchamps and Wahba, 2006). Fafchamps and Shilpi (2003) show that the distance to cities crucially determines wage opportunities and employment structure in Nepal and thus nonfarm employment (either wage or selfemployment) is concentrated in and around cities. Since road construction improves the access to (nonagricultural) labor markets or urban consumers, it increases wages and employment choices for rural residents. Some types of employment may be also made more possible by improved spatial linkages than others. For example, the study of Fujita and Muto (2007) on brand agriculture implies that the effect of spatial linkages depends on the type of differentiation of products. The connectivity to urban centers benefits laborer households more than farm (landed) households by improving the access to nonagricultural employment opportunities. Foster and Rosenzweig (2001) recently showed evidence from India that the landless prefer road construction as a local public investment choice because it improves the access to labor market, while the landed prefer investment in irrigation augmenting returns to land. The improvement of spatial connectivity also has implications on product markets, reducing transportation margins. Minten and Kyle (1999) showed that price variations are largely due to the transportation cost in the former Zaire. Interestingly, traders gain from bad road conditions with reduced purchase prices (increasing their profit). Therefore, spatial connectivity can potentially increase farmersincomes by reducing tradersprofit margin. A quite number of studies have estimated returns to infrastructure investment such as road construction under various assumptions but mostly at the aggregate level (Fan, et al, 2004; Binswanger, et al 1993). To analyze the dynamic effects on income growth at the household level, however, we must combine both household and spatial panel data over a long span of time with sufficiently large changes in infrastructure by household/village locations. In this paper, we try to capture the improvement of spatial connectivity by constructing the average intervillage road quality data (from the Indonesian village census) and combining this with the data on distance to economic centers categorized as subdistrict center, district center, and provincial center in Indonesia (from the village survey we conducted in 2007). Our main idea is that intervillage road quality determines the transportation means used in the local economy, therefore the average speed for resource mobility (including human), which affects the allocative efficiency in the local economy. The allocative efficiency is also affected by the distance to economic centers at different levels as these economic centers may offer dif-

3 ferent economic opportunities. Previous studies on spatial connectivity of rural households were limited in the sense that they perceived connectivity only as access to local towns or remoteness from growth centers, not being able to discuss the combination of both. But in actual policy choices, public investment planners face decisions on the allocation of resource among trunk roads (that lead to economic centers) and local roads. They also face the policy choice regarding the balance between spending on education and on roads. Therefore, this paper is intended to bridge the gap between academic studies and infrastructure planning. Empirical results show that the impacts of the improvement of quality of local road in the local area (positively correlated with transportation speed) on income growth and transition to nonagricultural activities depends on the distance to economic centers and household education. Education significantly increases the benefit from the spatial connectivity improvement, which is augmented by the distance from provincial center. Education and local road quality are complementary, increasing income growth. Therefore, whether the local connectivity improvement (measured by the average road quality) is propoor or not depends on village location and the initial householdlevel endowment. Data The data we use come from two sources. First, the main data come from village and household level surveys which we conducted in 2007 for 98 villages in 7 provinces (Lumpong, Central Java, East Java, NTB, South Sulawesi, North Sulawesi, and South Kalimantan) under the JBICs Study of Effects of Infrastructure on Millennium Development Goals in Indonesia (IMDG). The 2007 village survey captured the physical distance and time to various economic activity points such as market, station, and capital towns. Figure 1 shows locations of surveyed villages. The survey was designed to overlap with villages in the 1994/95 PATANAS survey conducted by ICASEPS to build household panel data. The 1994/95 PATANAS survey focused on agricultural production activities in 48 villages chosen from different agroclimatic zones in 7 provinces. In 2007, we revisited those villages to expand the scope of research as a general household survey under the IMDG survey. In the 2007 round, therefore, we added 51 new villages in the 7 provinces. In the revisited villages, we resampled 20 households per village from the 1994/95 sample and followed the split households. In the new villages, we sampled 24 households from two main hamlets in each village. Since one of the 48 villages in the 1994/95 PATANAS was not accessible for safety reasons in the 2007 survey (NTB province), we have the total of 98 villages that are available for various research objectives. In our panel analysis, we constructed household income panel data from 34 villages in 6 provinces (Lumpong, Central Java, East Java, NTB, South Sulawesi, North Sulawesi) using both the 2007 household and 1994/95 PATA- NAS surveys /95 PATANAS survey consists of two subsurveys. Income and production data are available from the second part, which contains 34 villages in 6 provinces excluding South Kalimantan. To merge the household panel data with spatial data on road quality constructed from PODES ( ), we use the information on subdistrict, district and province identification. In the analysis, we use subdistrict and districtlevel road quality variables to be interacted with household and villagelevel variables such as land owned and distance to district center. At this stage, we found that we cannot construct road quality data for 2 subdistricts in North Sulawesi as they have missing information in PODES.

4 Second, 1996 and 2006 PODES data were used to construct road quality data. PODES is a village census conducted by the Republic of Indonesia Central Bureau of Statistic. Details on are described in Section 3. Descriptive Analyses In this section we describe village census data: PODES with focus on transportation and road quality variables, and characterize changes in local road quality in the period of 1996 to The data cover all villages in the census years. For our research, we use 1996 and 2006 rounds as our household panel data were collected in 1995 and In the panel analysis, we take the difference between 1996 and 2007 to represent changes in the average road quality in the local economies. The PODES data have the information on major intervillage traffic. If the major traffic is on land, they ask about the type of widest road for this purposeasphalt/concrete/coneblock, hardened, soil, and others. Another question identifies whether 4wheel or more vehicles pass the road all year long. From the above information, it is possible to construct indicator variables for ( i ) major intervillage traffic land or not, ( ii ) type of widest roadasphalt/ concrete/coneblock or not, (iii) type of widest roadhardened or not, (iv) type of widest road soil or not, ( v ) type of widest roadothers or not, and (vi) 4w or more vehicle can pass the road all year longyes or not. We choose the measure ( ii ) to capture transportation speed in the local economy. The average is taken at the subdistrict, district and province levels in each round. z t ( j) mn( j) z m t #N( j) where z m t is the indicator variable which takes

5 the value of one if major intervillage traffic is on land and the road is constructed of asphalt/ concrete/coneblock (good quality) and zero otherwise (bad quality), N( j) is a set of villages within the is neighborhood, and #N( j) is the number of villages in N( j). Therefore, z t ( j) is the probability of having goodquality transportation, which is assumed to be positively correlated with the average transportation speed in the local economy. Table 1 shows the provincewise averages of asphalt road indicators in 1996 and To have comparability between the two years, we province Unit of observations is village Source : PODES 1996/2006 use 1996 provinces for villages which have changed province/district from 1996 to First, in both years, we observe interprovincial disparities in the average road quality. Second, the average proportion of asphalt intervillage roads has improved in many provinces. Table A1 in Annex shows tabulations of villages matched between 1996 and 2007 based on changes in intervillage road quality (asphalt or not). In many provinces, more villages have improved intervillage road quality rather than deteriorated although a large number of villages have no change in quality and there are a non negligible number of villages where road quality has been deteriorated. The reason for deterioration of road quality is not obvious from the data. Yet, it may be related to inadequate road maintenance or construction of new road with poor quality. Next, taking difference between the two rounds, we can see improvement and deterioration of road quality in local economies: z( j)z 1 ( j)z 0 ( j) Interestingly, we found that, in all regions, the changes are symmetrically distributed with either improvement or deterioration though the majority shows relatively small changes around zero (see Figures A1 in Annex). At the subdistrict level, improvement and deterioration coexist over the ten years in Indonesia, by which we can examine the impact of intervillage quality change on household income dynamics. Comparison of the road quality change (at the subdistrict level) between Java and nonjava regions showed that Java areas had experienced a faster improvement than outside Java. We assume that the physical distance has been constant throughout the period, so it is taken as predetermined. This information is important because we think the impact of spatial connectivity development on village economies

6 Distance (km) Distance (km) Province Village sub-district district province Province Village subdistrict district province Lampung Central Java East Java West Nusa Tenggara Source : IMDG South Kalimantan North Sulawesi South Sulawesi Mean

7 is not even, depending on the distance to main economic activity points. Table 2 shows distances to the centers in all 98 villages, using the 2007 village survey. In the analysis of household income dynamics, we use household panel data from two rounds conducted in 1995 and 2007 in 6 provinces as mentioned above. In both surveys, we collected detailed information on income generating activities. From each activity, we aggregated incomes to construct householdlevel income measure. To merge the income data with that of 1995, we aggregated incomes from original and split households using the 1995 household units. Some households split from the 1995 households (called original households), but it is important to aggregate incomes from both original and split households in 2007 to be comparable with the 1995 original households. The results were quite similar, which implies attrition (split) bias in our panel analysis was not large (See Figures A2a and A2b in Annex on percapita 3 income growth and change in nonagricultural income share). Table 3a shows descriptive statistics of key variables: number of household members aged 1564, household incomes, its growth, nonagricultural income shares, nonfarm selfemployment income shares, landholding size and 1995 household heads education in the panel sample. First, both nonagricultural and non farm selfemployment income shares increased in the period. Second, about 23.6 percent of the sample households are landless. Third, about 10 percept of the households had heads who completed high school or above. Lastly, growth of nominal household income is about 1.8. However, we have to note that regression analysis always includes location averages (dummies) which controls price changes specific to each location (village). Provincewise averages are compared in Tables 3b. First, nonagricultural income and non farm selfemployment income shares are high- Variable N Mean Std. Dev. Min Max Number of household members aged 1564 in Number of household members aged 1564 in Household income 2007 ( Rupiah) Household income 1995 ( Rupiah) Income growth Nonagricultural income share Nonagricultural income share Nonfarm selfemployment income share Nonfarm selfemployment income share Landholding size 1995 (ha) Landless indicator 1995 (1: landless) Heads years of schooling Head completed at least primary school 1995 (0: not completed) Head completed high school or above 1995 (0: not completed) Only those who are aged 1564 are considered to calculate per capita income growth.

8 Variable Lampong Central Java East Java NTB North Sulawesi South Sulawesi Number of household members aged 1564 in Number of household members aged 1564 in Household income 2007 (100,000 Rupiah) Household income 1995 (100,000 Rupiah) Income growth Nonagricultural income share Nonagricultural income share Nonfarm selfemployment income share Nonfarm selfemployment income share Landholding size 1995 (ha) Landless indicator 1995 (1: landless) Heads years of schooling Head completed at least primary school 1995 (0: not completed) Head completed high school or above 1995 (0: not completed) Source : IMDG 2007 and PATANAS 1994/95 Panel Data er in Java provinces than outside Java. Second, this does not necessarily imply higher income (or growth) in Java provinces. Third, landholding size is smaller in Java provinces than outside Java. It is easy to link diminishing roles of land and increase in nonagricultural activities in rural areas, but this does not mean higher income or its growth in our sample. Relationships to changes in local road quality are described in graphs below. To merge the household panel data with spatial data on road quality constructed from PO- DES ( ), we use the information on subdistrict, district and province identification. In the analysis, we use subdistrict and district level road quality variables to be interacted with household and villagelevel variables such as education, and owned and distance to district center. Figure 2a shows the relationships between change in the asphalt road proportion (at sub district level) and percapita income growth in our sample. Since price change and province level aggregate factors affect income growth (as well as the road quality change), we demeaned percapita income growth by controlling province effects to obtain the residuals. Therefore, the figure shows the effect of local road quality change on the residuals. Local road quality change and income growth are positively related, which motivates our analysis 4. Second, Figure 2b depicts the relationship between changes in the asphalt road proportion and nonagricultural income share. It clearly shows a positive association between the two changes. Though we face some identification issues in the estimation we conduct below, the above relationships back up our motivation. Next, we investigate the relationship between heads years of schooling and income growth or change in nonagricultural income share. In this exercise, we use observations (villages) Figures A3 (in Annex) disaggregate the sample in two groups by distance to province center. Figure A3a (Figure A3b) shows the relationship between percapita income growth residuals and road quality improvement in the villages with more than (less than) 100 km from provincial center. The positive relationship is clearer in villages with long distance from provincial center.

9 which experienced a positive change in the road quality in their subdistricts. Figures 3a and 3b show percapita income growth and nonagricultural income share respectively. Income growth is demeaned by village effects, so we observe intravillage variations using the residuals. Interestingly, as heads years of schooling increases, income growth stays intact up to around junior highschool completion, but it substantially increases from senior highschool completion or higher. There seems to be a threshold in schooling level, beyond which local road quality change and education jointly increases the impact on income growth. In contrast, Figures 3b shows a clear negative (monotonic) effect on change in nonagricultural income share. Less educated households (measured by the heads schooling) are likely to increase nonagricultural income (activity) share when road quality improves in their neighborhood. In this section we describe nonagricultural income opportunities in rural Indonesia, using the 2007 household survey data for 98 villages, with the focus on the nonfarm selfemployment and its linkage to the spatial connectivity

10 Total Distance to district center (km) Distance to province center (km) Mean share in household income by: Nonagricultural sector Selfemployment (nonfarm) 22% 21% 24% 21% 26% 22% 20% 19% Nonagriculture employment 22% 30% 21% 17% 27% 22% 19% 16% Agricultural sector Farm activities 36% 32% 38% 39% 28% 38% 43% 42% Agriculture employment 19% 18% 17% 23% 20% 19% 18% 23% Share of households with: At least one selfemployment activity 37% 32% 40% 37% 41% 37% 33% 33% At least one manufacturing activity 13% 11% 12% 16% 15% 10% 12% 18% Distribution of all households by distance 100% 30% 37% 33% 33% 28% 31% 9% Source : IMDG 2007 to economic centers. As it is shown in Table 4, the mean share of nonagricultural income in total household income is about 45%, and the share of nonagricultural income reduces as the distance from either district or province center increases. Within the nonagricultural income, which is composed of nonagricultural labor income and nonfarm selfemployment income in this analysis, nonagricultural labor income share declines by distance from economic centers. However, interestingly, the mean share of nonfarm selfemployment in household income does not necessarily decline. In fact, it is also not necessarily that the distance from district center reduces the share of households which engage in nonfarm selfemployment activity. In particular, this is the case for selfemployment activities which involve manufacturing or processing activities. For example, the share of households with at least one nonfarm selfemployment manufacturing activity is 11% among households living within 15 km from district center while the share is 16% among those far more than 30 km. The manufacturing activities account for nearly a half of selfemployment activities in the survey. The main products include processed food, such as dried fish and crackers, wood (or bamboo) products and garments (Table 5). Next, we illustrate the density (frequency) of households with selfemployment activities by distance from district centers, using graphs based on kernel density estimates 5, to obtain insights on how nonfarm selfemployment activities are linked to spatial connectivity to economic centers and what type of selfemployment activities are made possible by spatial linkages. In each graph, we compare the density pattern of households with at least one self employment manufacturing activity with the density pattern of household with other self employment activities. The density pattern of all households (either with selfemployment or not) is also presented as a reference (dot lines). We use Epanechnikov kernel function and a default bandwidth in the application. We also estimated with some alternative bandwidths but key messages presented in this paper are almost held the same.

11 Total Distance to district center (km) Distance to province center (km) Manufacturing Proceessed food 26.9% 25.1% 26.3% 29.1% 26.3% 25.9% 28.8% 26.3% Wood/bamboo products 8.2% 4.2% 3.8% 16.7% 10.5% 5.5% 3.4% 25.0% Cloths/textile 6.3% 8.7% 6.9% 3.5% 5.9% 4.8% 8.5% 5.3% Building materials 0.5% 0.7% 0.5% 0.3% 0.3% 0.7% 0.3% 1.3% Others 6.6% 7.7% 5.7% 6.6% 6.7% 10.0% 3.4% 5.3% Non manufacturing 51.6% 53.7% 56.7% 43.8% 50.4% 53.1% 55.6% 36.9% Total 100% 100% 100% 100% 100% 100% 100% 100% Source : IMDG 2007 We also generate graphs after separating households into two groups: one for better spatial connectivity to the economic centers and the other, and compare the results 6. Some graphs are presented below. Figures 4a and 4b show that selfemployment manufacturing activities exist more at distant (but not very far) places from district centers. This is clearer among households in provinces where the density of national and provincial roads is relatively high than among those in low density provinces. This implies that improving road networks beyond the districtlevel may enable manufacturing selfemployment ac- We used two types of spatial connectivity indicators. First, we calculated the national and provincial road density, in terms of road distance per area, as a proxy of road network for each of seven surveyed provinces, using data in JBIC study (2002) because development of road network within an economic region (e.g. province) as well as a route to economic centers is important for measuring spatial connectivity. (Table A2). The road density ranges from 0.04 in South Kalimantan province to 0.13 in North Sulawesi province. Second, we calculated speed indicators (km per minute) to reach district or provincial center, using data on time to go there by the most common mode of transportation (see Figure A4), as physical distance may have different implications in terms of the connectivity to economic centers, depending on various factors such as road and traffic conditions. In fact, speed indicators are not significantly correlated with physical distances in our data.

12 tivities to emerge at distant places from district centers (but not very far places). Although we need further investigation about why the occurrence of manufacturing activities is high around in villages in between 40 km to 55 km from district center, this may be related to better access to local resources (e.g. woods) 7 and the reasonable range of transportation time or cost (e.g. within two hours) to reach wider consumers including those in urban provincial centers. Figures 5a and 5b show similar density patterns to previous ones. That is, selfemployment manufacturing activities exist at distant (but not very far) places from district center. However, there is no clear difference in this pattern between households in villages where speed to access to a nearest district center is relatively high or those in low speed villages 8. This implies that the possibility for distant households to engage in nonfarm selfemployment activities may not be changed by the improvement of districtlevel roads via faster speed (reduced time) of transportation. Empirical Framework In the analysis we estimate the following equations on income growth and change in non agricultural income share, both first differenced between 1995 and 2007 to eliminate fixed effects. Both income growth and nonagricultural income share equations, after first differenced, are written as: y ij 11 z( j) 12 x 0 ijz( j) 21 d j z( j) 22 x 0 ijd j z( j) ij where y ij is income growth (or change in nonagricultural income share) for household i in village j, z( j) is change in the average road quality in the neighborhood of village j, d j is the distance to a center (to be discussed below), x 0 ij is household is land owned and education in the initial period, and ij is an error term. As mentioned, fixed effects are differenced out. We assume that the distance to economic activity center is predetermined, so taken as exog- The data shows that wood (including bamboo) related manufacturing activity accounts for over 25% in 4055 km to district center, while the main manufacturing activity in other areas is food processing activities. The high speed group includes villages where the speed (km/minutes) to access to a nearest district center is equal or more than 0.56 km in dry season and the low speed group includes other villages.

13 enous. Economic activity point can be subdistrict, district or province center. The interaction of z( j) and d j captures how the benefit from the spatial connectivity improvement varies with village location and distance from economic activity points. In the above specification, we also attempt to capture heterogeneous effects of the spatial development by the household initialstage asset holding and endowment. We use the information on landholding size and household heads education in The error term potentially consists of aggregate and householdspecific shocks: ij j i. To control provincespecific shocks, we can include province dummies. However, villagespecific shocks are correlated with local economic development, which is again correlated with dynamic change in the average road quality. Thus, E[ j z( j)]0. In the estimation below, therefore, we control villagelevel dynamic shocks in the first differenced specification. y ij 12 x 0 ijz( j) 22 x 0 ijd j z( j) village dummies ij This specification enables us to see intravillage variations in the response to the spatial connectivity development (as the village average is controlled). Villagespecific income shocks (affecting growth) are controlled by village dummies. We assume that the correlation between householdspecific shocks and the areawide spatial development is not important. Note that we use income aggregated from both original and split households in Therefore, our results will be robust to attrition bias potentially arising from endogenous household split dynamics. In the analysis, however, individual migration process is taken as exogenous, which may bias our estimates given that the migration process defines the denominator to calculate percapita income. Empirical Results In this section we summarize main results from the household analysis. In this analysis, we examine household income growth, changes in nonagriculture income share, and nonfarm selfemployment income share. In preliminary analyses, we found that subdistrict level road quality measure explains them better than districtlevel and provincelevel road quality measures, probably because it has enough variations in the sample and localized spatial connectivity development is important to opening access to wider economic activities (such as district and province center). To capture potential heterogeneous effects of the subdistrict average road quality improvement on income growth, we introduce some heterogeneity in the analysis: household heads education level and landownership in 1995 at householdlevel and distance to subdistrict, district and provincial centers at the village level 9. The main analytical point is to investigate the role of postprimary education and initial landholding in income growth when spatial connectivity is improving in the local neighborhood, and then to investigate the relationship with the connectivity with farther economic centers. We include village dummies to control villagespecific shocks containing price change specific to In our empirical setting with a small number of villages in each subdistrict, we cannot identify the effect of subdistrict level road quality change on householdlevel outcomes. As Figures 3 show, the effect seems to be positive in income growth and negative in nonagricultural income share. In preliminary analyses, we could not find significant effects with provincelevel dummies except a few cases. Therefore, we focus on intravillage distributional effects (with village dummies controlling price change and villagelevel shocks) in our parametric estimation.

14 village economy. In Table 6, Columns 1 and 2 use years of schooling completed, being interacted with distances to subdistrict, district and province centers. The results confirm that schooling effect is significantly positive (in the specification with the squared term). Interactions with distances are not significant. Column 3 uses the indicator which takes the value of one if head has completed high school or higher, and zero otherwise. Consistent with Figure 3a, it is a significantly positive effect. Interestingly, the effect increases as the distance from province center increases, and decreases as the distance from district center increases. Returns to schooling decrease if village is far from district center but the distance from provincial center significantly augments the returns. Thus, if village is near local center (district center) but the location of local economy is far from provincial center, the benefit from the spatial connectivity improvement becomes larger among relatively educated villagers. The above results suggest that local center in remote area is key. Marginal benefit from local road quality improvement is large in remote areas, probably because capital accumulation is at low level. However, our results show that district center is always important in local economy given localized economic interactions at Dependent variable: Percapita income growth Education variable: Years Years High school or higher Years High school or higher Change in average road quality Educ (1.16) (1.86) (1.80) (1.79) (1.93) Educ squared (1.08) (1.00) Educ Distance subdistrict (1.56) (1.62) (0.49) (1.00) (0.28) Educ Distance district (1.03) (0.66) (2.25) (0.84) (2.43) Educ Distance province (1.17) (0.93) (2.70) (1.04) (2.89) Land size (0.03) (0.04) Land Distance subdistrict (1.13) (0.97) Land Distance district (1.42) (1.57) Land Distance province (1.86) (2.07) Village dummies yes yes yes yes yes Rsquared Number of observations Numbers in parentheses are absolute t values, using robust standard errors with villagelevel clusters. Source : IMDG 2007 and PATANAS 1994/95 Panel Data

15 district level. There seems to be two important dimensions in their economic connectivity: links to local economy (district capital) and larger economic demand center (province capital). In the former, proximity to the center is always beneficial for the educated, but areas far from the latter (thus, districts far from province capital) are more likely to benefit from local road quality improvement. Regardless of interactions with distance, education always increases marginal benefits from local road quality improvement. Columns 4 and 5 include landholding size effects. Though landholding does not show significant effects on income growth, the exercise proves the robustness of our previous findings on schooling. Land is an important conventional input in agricultural production. But since the land is already used in 1995, its conversion into nonagricultural or financial resources always incurs opportunity costs too. In our findings on income dynamics, land does not matter in income growth and nonagricultural transition, which does not exclude its static contribution to agricultural production. Next we examine change in nonagricultural income share (Table 7). Columns 1 and 2 examine schooling effects on nonagricultural income share. Consistent with Figure 3b, we found that schooling decreases change in nonagricultural income share. That is, the (positive) effect of road quality is larger among uneducated households. Distance from subdistrict center dimin- Dependent variable: Nonagricultural income share Education variable: Years High school or higher Change in average road quality Years High school or higher Educ (2.33) (2.2) (2.44) (2.27) Educ Distance subdistrict (0.78) (5.45) (0.59) (5.66) Educ Distance district (0.77) (0.20) (0.90) (0.26) Educ Distance province (0.47) (0.03) (0.65) (0.10) Land size (1.25) (1.16) Land Distance subdistrict (1.22) (0.67) Land Distance district (1.22) (0.93) Land Distance province (0.05) (0.30) Village dummies yes yes yes yes Rsquared Number of observations Numbers in parentheses are absolute t values, using robust standard errors with villagelevel clusters. Source : IMDG 2007 and PATANAS 1994/95 Panel Data

16 ishes the above effect. Columns 3 and 4 include landholding sizes interacted with distances to economic centers. Landholding does not matter in the transition to nonagricultural activities. Education effects remain robust with landholding size. There are possibly three reasons for the negative effect of schooling on change in nonagricultural income share. First, the educated are more likely to have the nonagriculture income opportunities than the less educated at the initial stage, and therefore the local road quality improvement has a smaller marginal effect on the transition to the nonagriculture sector among the educated. Second, the more educated households also have more assets for agriculture production and thus the road quality improvement increases the productivity of their farm activities. Third, an individuallevel selectivity may cause the above result. At the individual level, the educated are more likely to move out of the households over time to get higher income opportunities in nonagricultural sectors. The comparison of completed schooling between current members and nonmembers shows higher average schooling among nonmembers. In the household with educated head, other members were also likely to be educated too. Therefore, if the above mentioned migration selection is important in the period of , an inverse correlation between schooling (at the household level) and observed nonagricultural transition is feasible. This is because edu- Dependent variable: Change in nonfarm selfemployment income share Education variable: High school or higher High school or higher Change in average road quality Educ (0.33) (0.01) Educ Distance subdistrict (1.95) (3.26) Educ Distance district (1.13) (1.20) Educ Distance province (4.61) (4.24) Land size Land Distance subdistrict Land Distance district Land Distance province (1.48) (2.01) (0.52) (1.63) Village dummies yes yes Rsquared Number of observations Numbers in parentheses are absolute t values, using robust standard errors with villagelevel clusters. Source : IMDG 2007 and PATANAS 1994/95 Panel Data

17 cated agents go out, and stayers are relatively less educated in the households. Table 8 shows results on change in nonfarm selfemployment income share. We use the same specifications adopted in the previous analyses. In Column 1, we found that the schooling effect critically depends on village location. Distance from provincial center reduces the schooling effect, while distance from subdistrict center increases the effect. This probably means that nonfarm business activities tend to pay off in areas close to economic centers with large demands (and its heterogeneity) such as provincial center. Large demand enables households to cover a relatively large setup cost. Distance effects are all negative in landholding effect, which is also consistent with the above finding. To sum, we show a summary table of our parameter estimate signs. First, interesting results are concentrated in education effects. In general, land does not matter in the dynamics of household income and nonagricultural transition. Second, while education augments the impact of road quality improvement (spatial connectivity) on percapita income growth, it decreases the impact on nonagricultural transition. Third, a similar contrast is also observed in the role of distance to different economic centers. In transition to nonagricultural activities among educated households, the marginal impact of local road quality improvement are large in locations distant from local economic centers (subdistrict capital), but in income dynamics the impact is large in villages far from provincial capital. Dependent variable: Percapita income growth Nonagricultural income share Nonfarm selfemployment income share Education variable: High school or higher High school or higher High school or higher High school or higher High school or higher High school or higher Change in average road quality Educ + + Educ Distance subdistrict Educ Distance district Educ Distance province Land size N/A N/A N/A Land Distance subdistrict N/A N/A N/A Land Distance district N/A N/A N/A Land Distance province N/A N/A N/A Village dummies yes yes yes yes yes yes Rsquared Number of observations One sign : significant at 10% level; two signs: at 5%level; three signs: at 1% level Source : IMDG 2007 and PATANAS 1994/95 Panel Data

18 In our definition, nonagricultural activities only cover those done by current household members. This excludes nonmembers who work in locations distant from their villages (not able to commute from their villages). Therefore, it is still possible that we are missing migrationlinked nonagricultural transition. Instead, income growth includes agriculturebased growth, which for example includes improved marketing of agricultural products (e.g., vegetables). In this activity, connecting to larger demand centers seems to be a driving force. Policy Discussion This paper is intended to bridge the gap between academic studies and infrastructure planning. Previous academic studies on spatial connectivity of rural households were limited in the sense that they perceived connectivity only as access to local towns or remoteness from growth centers, not being able to discuss the combination of both. But in actual policy choices, public investment planners face decisions on the allocation of resource among trunk roads (that lead to economic centers) and local roads. Public investment planners also face the policy choice regarding the balance between spending on education and on roads. The analyses described above suggest that the more educated households can increase income with better spatial connectivity at local level. Better local road quality may also improve the access for remote villages to trunk roads and thus help the more educated engage in better job/business opportunities at district capital (local economy) or province capital (larger economic center). However, proximity to district center augments the effect to income growth while distance from provincial center augments it. Although we cannot include in the empirical analysis due to data limitation, this difference may be due to the market space as well as the value added of different income generating activities. First, there exists income generating activities focusing on the market with district capital as the local economic center. These may include activities such as food processing with low value added (such as dried fish or chips/ crackers) and marketing staple food. In this case, proximity to the economic center is a key as it reduces transport related transaction cost. However, there are other types of activities with wider market area, especially catering to urban economic centers such as provincial center. These may include higher value added goods sold in large urban markets such as bamboo or wood products. Another example can be high quality vegetables for the urban market. In such case, the added value can cover the transaction cost due to transportation and thus distance from provincial center is not an obstacle, provided that it is connected to economic centers. Better road connectivity to provincial center due to local road improvement may give remote villages the chance to market such value added products. In the former case, it can be suggested that improving the trunk roads connecting to closer district centers is important alongside with the improvement of local roads that provide access to such trunk roads. In the latter, it is important to develop the network of the trunk roads to secure connectivity to distant economic centers, such as provincial capital, alongside with the improvement of local roads. Poverty Reduction Strategies (PRSs) adopted by low income countries especially those in Africa are entering a second stage, becoming more growth oriented. Compared to the previous generation of PRSs emphasizing budget allocation to primary education and health, the current generation focuses on growth strategies. Yet, little is known on the type of public invest-

19 ment combination that induces growth. The analyses of this paper suggest that investing simultaneously in spatial connection of local neighborhoods as well as in connecting to distant economic centers pays off. This paper also suggests that investing in both higher education (high school and above) and roads is important. Although the actual PRSs should be country driven and country specific, such findings can add value to the next generation of growth oriented PRSs. Conclusion This paper examined the impact of spatial connectivity development, on household income growth and nonagriculture income share, combining household panel data and village census in Indonesia. Empirical results show that the impacts of the improvement of road quality in the local area (positively correlated with transportation speed) on income growth and transition to nonagricultural activities depends on the distance to economic centers and household education. In particular, postprimary education significantly increases the benefit from the local connectivity improvement in remote areas. Post primary education and local road quality are complementary, increasing income growth. Therefore, the effectiveness of local connectivity improvement (measured by household income growth) depends on village remoteness and the initial householdlevel endowment. References Andrew Foster and Mark Rosenzweig, 2001, Democratization, decentralization and the distribution of local public goods in a poor rural economy, Manuscript, Brown University. Fujita Masahisa and Muto Megumi, 2007, Brand ngyo niyori taysei no yutakana tojkoku kaihatsu wokkan keizaigaku no shiten kara (Development in developing countries employing brand agriculture from the view of spatial economics), Journal of JBIC Institute No. 33. (in Japanese) Hans P. Binswanger Shahidur R. Khandker, and Mark R. Rosenzweig, 1993, How infrastructure and financial institutions affect agricultural output and investment in India, Journal of Development Economics, 41: JBIC, 2004, Sector study in road sector in Indonesia,JBIC Sector Study Series 2003No. 2. Marcel Fafchamps and Forhad Shilpi, 2003, Spatial division of labor in Nepal, Journal of Development Studies, 39: Marcel Fafchamps and Forhad Shilpi, 2005, Cities and spacialization: Evidence from South Asia, Economic Journal, 115: Marcel Fafchamps and Jackline Wahba, 2006, Child labor, urban proximity and household composition, Journal of Development Economics, 79: Minten and Kyle, 1999, The effect of distance and road quality on food collection, marketing margins, and traderswages: evidence from the former Zaire, Journal of Development Economics, 60: Shenggen Fan, Linxiu Zhang and Xiaobo Zhang, 2004, Reforms, investment, and poverty in rural China, Economic Development and Cultural Change, 52:

20

21 Province name No change Remain Remain good bad Number of villages Total Proportion of villages in each province No change Remain Remain good bad Deteriorated Improved Deteriorated Improved Difference (Improved) (Deterorated) Jawa Barat , % 38.5% 16.2% 9.0% 7.2% Lampung % 11.5% 10.2% 6.7% 3.5% Maluku % 46.0% 12.0% 9.2% 2.8% Jambi % 16.8% 11.0% 8.4% 2.6% South Kalimantan % 11.0% 9.8% 8.2% 1.6% East Java 1, , % 21.5% 13.7% 12.3% 1.4% Aceh 989 1, , % 45.0% 16.3% 15.3% 0.9% Kalimantan Timur % 0.5% 1.3% 1.6% 0.3% Bali 1,277 1, , % 38.0% 11.4% 12.6% 1.2% Sulawesi Tengah % 19.9% 11.3% 13.1% 1.8% Central Java % 0.0% 0.0% 2.6% 2.6% Riau , % 33.5% 7.8% 10.6% 2.8% West Nusa Tenggara % 54.0% 8.0% 11.1% 3.1% Sumatra Barat % 34.4% 9.3% 13.0% 3.7% Sumatra Selatan % 60.0% 2.0% 6.1% 4.0% Irian Jaya 1, , % 29.0% 7.1% 11.7% 4.7% Nusa Tenggara Timur % 78.6% 2.6% 8.4% 5.8% North Sulawesi , % 32.2% 8.3% 14.6% 6.3% Sumatera Utra % 53.5% 3.6% 10.4% 6.8% Bengkulu % 12.8% 2.8% 9.7% 6.9% Sulawesi Tenggara , % 34.8% 6.0% 13.1% 7.1% South Sulawesi % 68.6% 2.5% 10.0% 7.5% DKI Jakarta % 19.5% 9.1% 17.5% 8.4% Kalimantan Barat 4,379 1, ,441 7, % 17.3% 8.7% 18.3% 9.6% DI Yogyakarta , % 51.7% 5.9% 16.5% 10.6% Kalimantan Tengah 3,653 1, ,746 7, % 22.1% 10.1% 21.9% 11.8% Total 20,044 13,550 4,305 6,594 44, % 30.5% 9.7% 14.8% 5.1% Source : PODES 1996/2006