Voting with Their Feet? Access to Infrastructure and Migration in Nepal

Using bilateral migration flow data from the 2010 population census of Nepal, this paper provides evidence on the importance of public infrastructure and services in determining migration flows. The empirical specification, based on a generalized nested logit model, corrects for the non-random selection of migrants. The results show that migrants prefer areas that are nearer to paved roads and have better access to electricity. Apart from electricity's impact on income and through income on migration, the econometric results indicate that migrants attach substantial amenity value to access to electricity. These findings have important implications for the placement of basic infrastructure projects and the way benefits from these projects are evaluated.


Introduction
Do migrants respond to di¤erences in access to public goods and services in addition to income prospects of potential destinations? The income di¤erence between the origin and the destination is the primary factor driving migration in the existing literature on migration (Greenwood (1975) and Borjas (1994), Lall, Selod and Shalizi (2006)). Along with income, recent literature has also highlighted the importance of migration costs as well as migrants' networks in determining migration ‡ows. How provision of public goods and services may in ‡uence migration in poorer developing countries remains sparsely studied. This issue is however important in these countries where provision of public goods varies widely across areas. In a Tiebout (1956) sorting model, such disparity in the provision of public goods such as roads, electricity, schools, hospitals, etc. should induce people to "vote with their feet" and to migrate to areas with better access to these infrastructures and services. 1 From a policy perspective, it is important to know how migration responds to the provision of public goods in developing countries for a number of reasons. First, regions within a typical developing country are usually characterized by stark di¤erences in poverty and welfare. Households with poorer attributes such as low levels of education, skills and assets are frequently observed to live in poor areas that are characterized by lack of public infrastructure and services (Shilpi(2011), Dudwick et al. (2011), World Development Report (2009), Kanbur and Venables (2005), Jalan and Ravallion (2002), Ravallion and Wodon (1999)). If migrants do respond to income as well as provision of public infrastructure and services, then migration can act as a powerful instrument in mitigating regional di¤erences in welfare. Second, di¤erential costs of provision of infrastructure and services along with a hard budget constraint often force governments in developing countries to prioritize placement of these public goods. If people do migrate to gain better access to public goods, then the government may be able to rely more on cost considerations to prioritize their placement. Finally, migration in response to public goods and services also has important implications for the way the bene…ts of public investment are evaluated. A typical evaluation strategy 1 of relying on variations in key outcomes such as income or household expenditure across areas with and without public goods would seriously underestimate the bene…t of the project. This is because migration in response to a new public good reduces the di¤erences in these outcomes across areas. Using census data from Nepal, this paper provides evidence on the extent to which access to public goods and services in ‡uences bilateral migration ‡ow across areas.
The determinants of bilateral migration have been analyzed mostly in the context of international migration (Grogger and Hanson (2011), Ortega and Peri (2013)) and inter-regional migration (Ghatak, Mulhern and Watson (2008), Andrienko and Guriev (2004)). 2 This literature however focuses primarily on income and migration costs as determinants of migration ‡ow. A recent literature examines how migrants' choice of destination is in ‡uenced by locational attributes including the state of public infrastructure and services. For a relatively richer developing country -Brazil -Lall, Timmins and Yue (2009) …nd that poor migrants are willing to accept lower wages to achieve access to better services while richer migrants are in ‡uenced only by income di¤erences. 3 Fafchamps and Shilpi (2013) …nd a statistically signi…cant and numerically large e¤ect of access to paved roads on migrants'destination choice in Nepal: migrants prefer a destination that is closer to a paved road. While contributing to this literature, the analysis in this paper di¤ers from the above papers in a number of ways. Instead of focusing on the destination choice of individual migrants, we analyze bilateral migration ‡ows across multiple sources and destinations. Our empirical speci…cation is derived from a model of utility maximization by the migrants proposed by Ortega and Peri (2013) and Grogger and Hanson (2011). We consider a generalized nested logit model where migrants …rst decide whether to migrate and then decide among the potential destinations. The advantage of this approach is that the resulting empirical speci…cation includes a correction term for the unobserved heterogeneity between migrants and non-migrants. The above mentioned papers (Fafchamps and Shilpi (2013), Lall, Timmins and Yue (2009)) in contrast side-stepped the issue of migrants'non-random selection by focusing on the choice of destination conditional on migrating. More importantly, we make a distinction between the productivity and amenity values of basic infrastructure and services. For instance, access to electricity allows …rms to automate production, shifting the production possibility frontier ("productivity e¤ect"). It also helps households to carry out essential chores e¢ ciently and to enjoy leisure more fully ("amenity e¤ect"). We develop a strategy to uncover the amenity values of infrastructure and services. The strategy relies on a two-stage estimation procedure in which a canonical migration model -ignoring access to infrastructure and services -is …tted at the …rst stage. The …rst stage estimation thus allows income to capture the productivity e¤ect of the public goods. To the extent these goods are targeted to more productive areas, income in the …rst stage estimation picks up that placement e¤ect also. In the second stage, the residual from the …rst stage is regressed on measures of access to infrastructure and services. By construction, this strategy provides conservative estimates of amenity values of public goods.
The empirical analysis of this paper utilizes the detailed migration information from the population census 2010 of Nepal. Due to the mountainous terrain of the country and limited agricultural potential in many areas, migration is an important livelihood strategy for the Nepalese people. The rough terrain makes the provision of basic infrastructure very di¢ cult with the outcome that large parts of the country are not well served by transport infrastructure. Geographical coverage of electri…cation remains rather low, serving only a third of rural households. In terms of access to infrastructure and services as well as stage of economic development, Nepal is comparable to many Sub-Saharan African countries. The large geographical variations in access to basic public goods along with vibrant migration ‡ows make Nepal particularly suitable for our study.
When a standard migration model is …tted, our empirical results con…rm the common …ndings of the migration literature that income and distance between source and destination are the two most important determinants of bilateral migration ‡ow. Consistent with the …ndings of Fafchamps and Shilpi (2013), we …nd that when measures of access to basic public goods are added as regressors, the magnitude of the income coe¢ cient declines substantially though it still remains statistically signi…cant. This result con…rms 3 that the income coe¢ cient in a standard migration model is likely to be biased upward. Our results show that access to electricity and paved roads are important determinants of migration: migrants prefer areas with better access to electricity and paved roads. The results from the two-stage estimation procedures indicate that migrants attach substantial amenity value to access to electricity as well. Moreover, we …nd that migrants of di¤erent skill levels (primary, secondary and tertiary education level) attach similar amenity values to access to electricity. Thus better access to electricity attracts migrants not only because it brightens their income prospects but also because it o¤ers better quality of life to them.
The rest of the paper is organized as follows. The conceptual framework and empirical speci…cation are presented in Section 2. Section 3 discusses the data. Section 4, organized in subsections, presents the empirical results. Section 5 concludes the paper.

The Model
We start from a simple model of migration where an individual makes a utility maximizing migration decision among multiple destinations within the country. Individual h in her place of residence s decides whether to stay at s or to migrate to any of i 2 I = f1; :::; N g: Let utility of individual h in location i be denoted U h i . Following the literature, we assume that utility U h i is a function of the income y h i (or consumption) that the individual can achieve in location i, of the prices p i he or she faces, and a vector of location-speci…c amenities A i (Bayoh, Irwin and Haab, 2006). The utility from migrating to a given destination i depends on the migrant's utility from income and amenities suitably adjusted for prices and on the costs C h si of moving from s to i. Following Grogger and Hanson (2008) and Ortega and Peri (2009), we make a distinction between factors that are shared by all migrants from the same origin and to the same destination, and individual speci…c factors. The utility in destination i can be expressed as: where si is an origin-destination speci…c term shared by all individuals migrating from s to i and h si is the individual migrant speci…c term. u i (y i ; A i ; p i ) is the expected utility of individual h in destination i.
The expected permanent income of individual h in destination i is the average income y i : In the empirical estimation, we allow di¤erences in incomes for workers of di¤erent skill levels. The expected utility in the destination depends also on the services and amenities available there along with the cost of living. This is important particularly for internal migration where individuals and households may move not only to capture income gain but also to avail themselves of better services and amenities -for instance better schools or health services -at destination. Similarly, C si is the average cost of migration from s to i. The cost term C si captures the physical distance between origin and destination. It also re ‡ects costs incurred by individuals due to social distances (e.g. cultural, ethnic and language di¤erences) between the origin and destination.
We assume that u is an increasing function of y i ; and A i ; and a decreasing function of p i : We assume that g is an increasing function of C si : Following Grogger and Hanson (2008), we assume that both u and g are approximately linear functions. In the case of no migration, the average expected utility is: where ; and are positive constants. The utility from locating in i can be expressed as: where > 0 is a parameter. The individual speci…c term h si denotes the idiosyncractic parts of the utility and cost associated with migration by individual h. There is now substantial evidence that migrants may be substantially di¤erent from non-migrants in terms of their ability, risk aversion and preferences.
Following Ortega and Peri (2009) we assume that: where " si is iid following an (Weibul) extreme value distribution. i is an individual speci…c term that a¤ects migrants only and its distribution is assumed to depend on 2 [0; 1). Given that " si has an extreme value distribution, then h si has also an extreme value distribution (Cardell (1991). The migrant speci…c term i does not depend on destination and can be thought of capturing the di¤erences in preferences for migration.
Ortega and Peri (2009) show that utility maximization under the distributional assumptions leads to the following condition: Subsitutiting for shares and solving for the logarithm of migration ‡ow (ln m si ), equation (4) can be re-written as: where si is the zero-mean measurement error, s is the origin …xed e¤ects and 1 = 1 ; 1 = 1 ; 1 = 1 ; and 1 = 1 . In the standard logit formulation ( = 0); the …xed e¤ects account for share of the stayers in the population along with income, amenity and prices at the origin [ s = ln(n s When migrants di¤er systematically from non-migrants in preference and ability ( 6 = 0);the …xed e¤ects include a correction term ( 1 ln m si ) for the average unobserved heterogeneity between migrants and non-migrants as well.
We estimate the speci…cation in equation (5) for bilateral gross migration ‡ows among districts in Nepal.
Following Grogger and Hanson (2011), we analyze sorting of migrants across destinations. Speci…cally, we analyze the variations in the skill mix of migrants to di¤erent destinations. We de…ne three groups of migrants in terms of their education level: those with (i) less than primary education, (ii) education between primary and secondary levels and (iii) above secondary level.
where j represents the education levels of migrants. The speci…cations in equations (5) and (6) are based on a linear utility and migration cost functions. A linear formulation can be interpreted as monetary income and cost whereas a log-linear speci…cation would imply as log income and time cost (Ortega and Peri (2013)). We performed estimation using both linear and log linear speci…cations.
Equations (5) -(6) are the basis of our main empirical estimation. A number of things are worth noting in the estimation of equations (5)(6). First, when su¢ cient numbers of migrants come to a destination, it is expected to have general equilibrium e¤ects on wages, incomes and access to services and amenities. 4 This would generate a potential endogeneity bias due to the fact that income and amenities in a destination resulted in part from the decision of many migrants to locate there. To eliminate this bias, we use lagged explanatory variables. More precisely, let T be the period for which we have information on explanatory variables and T + t the period at which we observe migrants. Migrants are de…ned as those who migrated 4 The e¤ect could be negative -e.g., congestion -or positive -e.g., agglomeration externalities. 7 between T and T +t whereas explanatory variables come from period T: Limiting the set of migrants in this fashion ensures that migration decisions are based only on information available prior to migration. Second, bilateral migration ‡ows between districts are not always positive. While our main estimation focused on districts with positive migration ‡ows, we also checked the robustness of our results for the sample which included districts with no migration ‡ows. We weight observations by destination population which corrects for potential heteroskedasticity of measurement error. The standard errors are clustered by destination districts to account for within (destination) district correlation of errors.

Empirical Speci…cation
The basic empirical speci…cations estimated from the data augment equations (5) and (6) with additional explanatory variables, leading to the following estimating equations: where Z i is a vector of locational attributes of destination i. The Z i vector includes controls for social proximity between source and destination in terms of language, religion and ethnicity. Following standard practice in the literature, we also include a measure of the unemployment rate as a control.
Suppose 1 = j 1 = 0, then equations (7 and 8) have the speci…cations that are comparable to speci…cations derived from the standard model of determination of migration ‡ows when migrants'preferences for better access to public goods and services are ignored. For simplicity, suppose, y i and A i are uncorrelated with rest of the explanatory variables in the above equations and ( 1 6 = 0; j 1 6 = 0):We estimate equation (7) ignoring A i : The estimated coe¢ cient of income in this case is: 8 where 1 is the estimated income coe¢ cient when A i is ignored and^ 1 and^ 1 are the estimated income and amenity/public goods coe¢ cients from the full speci…cation (equation 7) and is the correlation between y i and A i : Since income tends to be higher in areas with better public goods, > 0: The income coe¢ cient ( 1 ) thus overestimates the in ‡uence of income di¤erences on migration ‡ow (^ 1 ) when migrant's preference for public goods is ignored.
The positive correlation between income and public goods means that part of the preference for public goods is due to a preference for higher income. Some of the basic public goods such as roads and electricity have not only direct productivity and hence income e¤ect but also amenity values as they make life easier for households. To explore the amenity value of these goods and services, we utilize a two-stage procedure.
At the …rst stage, we …t a standard migration model ignoring public goods, which allows income to pick up the productivity e¤ect of public goods and services. At the second stage, the residual from the …rst stage is regressed on the explanatory variables representing access to public goods and services. Similar to equation (9), it follows that: where 1 is the estimate of 1 from the second stage regression. 1 thus provides an estimate of the amenity value of public goods and services to migrants.

Data
The empirical analysis in this paper draws data from various sources: information is also available on gender, age, education, religion, language spoken, ethnicity and motive for migration. This rich data-set is used to de…ne the gross bilateral migration ‡ows across districts. 5 5 Nepal is divided into 75 districts. who migrated during the last 6 years, 54 percent moved for work reasons. We estimate the migration ‡ows among districts using the census data and appropriate population weights. We de…ne two types of migrants: work migrants who moved to seek employment, and all migrants including work migrants as well as those who moved for non-work related reasons. All estimations are carried out for both of these samples. Figure 1 shows the geographical distribution of migrants in terms of district of residence and origin.
As apparent from Figure 1, a small number of destination districts have a high proportion of migrants. In contrast, districts of origin are distributed widely across the country. This re ‡ects the fact that much of the migration is from rural areas to towns and cities. Indeed more than 90 percent of the migrants come from rural areas, and more than half of them migrated to an urban area. The same migration pattern is observed for work migrants.
While the census provided information about migration, it did not collect any information on income, prices and access to services and amenities. We utilize a nationally representative survey of householdsthe Nepal Living Standard Survey 2002/3 -to derive these explanatory variables. To estimate the average income level in a district i, we ran a regression of the following speci…cation: In the empirical analysis, migration costs are captured by geographical and social distances. For geographical distance between districts, we utilize the arc distance between the district of origin and each possible district of destination, computed from the average longitude and latitude of each district. 6 We expect the cost and risk of migration to increase with physical distance. Social distance captures the e¤ect of migration networks which are found to be important in determining migration ‡ow ( Munshi (1993), Beaman (2012)). Social distance is measured by the index of ethno-linguistic fractionalization (ELF). The ELF index measures the probability that two individuals taken at random belong to the same ethnic or linguistic group. We estimated ELF for each district using the method suggested by Alesina and La Ferrara (2005). The ELF measures are de…ned for religious, linguistic and ethnic-caste groups using data from the 2000 population census. We computed the district level unemployment rate from the census 2000 data.
Instead of using the share of households with electricity, we construct a measure of electricity connection which does not depend directly on household income. We compute the share of wards -the smallest administrative unit in Nepal -that had electricity connection among all wards in a district using census 6 The average longitude and latitude of a district are obtained as a weighted average of the longitude and latitude of all the VDC's in the district, where the population of each VDC serves as weight.

11
2000 data. 7 This de…nition of access to electricity avoids the correlation with income that would have resulted from the ability of households with higher incomes to get electricity connection had the access variable been de…ned at the household level. As a measure of access to markets and other services (schools, To control for price, we use price of rice. Rice is the most commonly consumed food item in Nepal and thus can be taken as a proxy for the price of common household goods. The NLSS 2002/3 collected information on the quantity and price paid for rice by individual households. We use these data to compute a unit price per kg. We construct a measure of housing cost using data from the NLSS 2002/3 survey which contained a separate section on housing. The survey collected information on hypothetical and actual house rental values of each household together with house characteristics such as square footage, number and type of rooms, quality of materials, and the availability of various utilities. We use these data to construct a hedonic index of housing costs for each district. Let r k i be the house rental price paid (or estimated) by household h in district i and let x h i denote a vector of house characteristics. We estimate a regression of the form: where estimate of b a i provides a measure the housing cost premium in each district i. In the regression -omitted for the sake of brevity -many of the house characteristics are signi…cant with the expected sign, e.g., larger, better built houses with better in-house amenities are worth more. District price di¤erentials are large and jointly signi…cant. Since the dependent variable is in log form, b a i measures the housing cost premium in each district. Table 1 reports the summary statistics for the dependent and explanatory variables. On average about 7 There are more than 35 thousand wards in 75 districts in our census long form sample.
12 45 people migrated between a source-destination pairs. The number of work migrants is smaller -about 19 people. Migration appears to be concentrated at the two ends of the skill distribution: both unskilled (up to primary education) and skilled (above secondary education) tend to migrate at a higher propensity relative to semi-skilled workers (above primary but up to secondary education). This is true for work migration also.
This pattern is consistent with the pattern observed for international migration into OECD countries. The propensity to migrate into OECD countries is lower at the semi-skilled level (see Table 1 Table 1.

Empirical Results
The initial set of regression results using the speci…cation in equation (7) are reported in Table 2. All regressions reported in this paper included birth district …xed e¤ects. All regressions are also weighted using destination district population, and all standard errors are clustered at the destination district level to account for any within district error correlations.

Determinants of Bilateral Migration Flow
We start with the simplest speci…cation which corresponds to the standard speci…cation estimated for bilateral migration ‡ow particularly in the context of international migration (Grogger and Hanson (2011), Ortega and Peri (2013)). The estimation is carried out for two samples: all migrants, and work migrants.
The results for this speci…cation are reported in columns 1 and 4 for the all migrant and work migrant samples, respectively. Consistent with the overwhelming evidence from the migration literature, migration ‡ow appears to be associated positively with income at the destination relative to that at source. The income coe¢ cients (columns 1 and 4) are quite precisely estimated. The estimated income coe¢ cient for the 13 all migrants sample is slightly larger in magnitude than that for the work migrant sample but the hypothesis that the two income coe¢ cients are equal cannot be rejected even at the 20 percent signi…cance level. We Among the other explanatory variables, we …nd that unemployment rate has the expected negative sign when controls for access to paved road and electricity, and housing price premium are added to the regression. The coe¢ cient of rice price also becomes statistically signi…cant though with a positive sign.
Rice price is higher in urban areas compared with rural areas where it is grown because of transportation cost. Rice price thus captures the fact that rural to urban migration is the predominant direction of migration in Nepal. Finally, coe¢ cient estimates are statistically indi¤erent between the two samples. For the rest of the paper, we thus limit our discussion to results from the full sample. In the next sub-section, 15 we explore if the results are di¤erent for migrants of di¤erent skills.

Determinants of Migration for Di¤erent Skill Levels
The determinants of migration ‡ow may be di¤erent for people of di¤erent skills. To the extent migrants with higher education come from relatively well-o¤ families, they may face lesser credit constraint in …nancing their migration including the time spell during job search. On the other hand, poorer and unskilled migrants may be pushed out of their source due to adverse shocks and hence their migration may be less sensitive to income di¤erences. To explore these possibilities, we divide migrants into three groups in terms of their education level. Skilled migrants are those with higher than secondary education, and unskilled with primary or less education while semi-skilled belong to the middle group. We report the estimation results for the regression speci…cations in columns (1) and (3) of Table 2. The regression results are reported in Table 3.
The overall results for all three skill groups are consistent with those reported in Table 2. Some patterns are however worth noting. Income di¤erences between the source and destination seems to have relatively smaller in ‡uence on unskilled migrants compared with semi-and skilled migrants, though income coe¢ cients are all positive and statistically signi…cant. The estimates of distance coe¢ cients on other hand display the opposite pattern: they are larger in absolute magnitudes for unskilled and semi skilled migrants compared with skilled migrants. This overall pattern is consistent with the expectations that many more of the unskilled migrants are push migrants and that because of credit constraint, they tend to migrate closer to their origin. Religious diversity -a factor that may relate inversely to migrants'social networkmatters much less for the skilled migrants.
Access to paved roads and electricity have statistically signi…cant coe¢ cients in all three samples. The when access to public goods and housing prices were ignored. This again con…rms that income and these locational attributes are signi…cantly and positively correlated. To the extent access to paved roads and electricity contributes to higher income, their respective coe¢ cients capture not only their amenity value but also their productivity e¤ect re ‡ected in higher income. In the following sub-section, we attempt to disentangle their amenity value.

Migration and Amenity Value of Public Goods and Services
To estimate the amenity value of public goods, we use a two-stage procedure. At the …rst stage, we estimate a standard migration model ignoring the di¤erences in the provision of electricity or paved roads across areas. This speci…cation thus corresponds to that in column 1 in Table 2, and columns 1, 3 and 5 in Table 3 The estimates for access to electricity suggest that migrants do attach amenity value to it. Even after conditioning on income, migration ‡ows are greater to areas which have better access to electricity. The results in Table 4 also suggest no substantial variations in the way migrants of di¤erent skill types value access to electricity as an amenity. The estimates of coe¢ cients of access to electricity in Tables 2 and 3 fall within the interval of [ 2.5-3.12]. The estimates in Table 4 are much smaller in magnitude -roughly about 40 percent of magnitude of estimates in Tables 2 and 3. In other words, of all the migration that happens in response to access to electricity, 40 percent of those is perhaps due to electricity's amenity value.
The estimates in Tables 2 and 3 suggested strong and negative association between bilateral migration ‡ows and travel time to paved road, the estimates in Table 4 show absence of a signi…cant association between these two variables. The strong and negative association between income and geographic isolation (measured here by travel time to paved road) is well noted in the case of Nepal (Fafchamps and Shilpi (2008) and (2013)). The lack of signi…cance of travel time to paved road in the second stage does not necessarily imply that migrants do not attach any amenity value to access to paved road. Rather it suggests that the correlations of travel time to paved road with income and with access to electricity are strong and that given those correlations, it is not possible to disentangle the productivity and amenity value of paved roads.

Robustness Checks
We perform a number of robustness checks. These checks are conducted for all di¤erent samples. To avoid clutter, we report the estimates of the coe¢ cients of access to electricity, paved road and housing price premium. We also report estimates from two regressions: a full model where all variables are introduced simultaneously; and the estimates from the second stage regression where …rst stage regression did not include any of the three variables of our interest. The full model thus corresponds to speci…cations whose results are reported in column 3 of Table 2 and columns 2, 4, and 6 of Table 3. The conditional estimates from second stage correspond to results reported in the even numbered columns of Table 4. We report the results for all migrants in Table 5.
The regression results reported so far come from speci…cations where income and distance variables are measured in levels. In most migration studies, these variables are often introduced in the logarithmic form.
The logarithmic form would imply a log-linear utility function which -according to Grogger  In the main regressions, we focused on the sample of all districts with positive migration ‡ows. In the next robustness check, we included all districts including those with zero migration ‡ow. The results are shown in the …nal two columns of Table 5. The results for this expanded sample is nearly indistinguishable to those reported in Tables 2 and 4.
We repeated the robustness checks for migrants of di¤erent skill levels. The results are similar to those reported in Table 5. In the upper panel of Table 6, we report the results when stock of past migrants is included as an additional regressor. The lower panel reports the results with log of 1991 population as an additional regressor. The results are comparable to those for the full sample. Though the magnitudes of coe¢ cients of access to electricity are somewhat smaller compared with those in Table 4, they are all numerically and statistically signi…cant. As before, we …nd that coe¢ cients of travel time to paved road 20 become statistically insigni…cant when population is added as a regressor. In all other cases, migration ‡ow seems to respond signi…cantly with access to paved road in the full regression.
It is worth noting that population of a district captures the relative degree of urbanization as well: districts with larger urban share also have higher population. Since urban areas di¤er distinctly from rural areas in terms of income and access to public goods, introduction of population in the regression leads to a substantial decline in the magnitudes of coe¢ cients of these variables. Thus inclusion of population as a regressor is likely to bias the estimates of income and public goods coe¢ cients downward. Same argument can be made about stock of past migrants. Our preferred speci…cations thus exclude these two variables.

Economic Signi…cance
The explanatory variables in the regressions are measured in di¤erent units and thus it is di¢ cult to compare the magnitudes of coe¢ cients of di¤erent variables. To provide an idea about the relative importance of di¤erent factors in determining migration ‡ows, we provide the estimates of the elasticities based on the estimated coe¢ cients which are reported in Table 7.
We computed elasticities for both the full model which included all explanatory variables simultaneously and the two stage procedure which excluded access to paved road and electricity and housing price premium from the …rst stage regression. In both models, distance between the source and destination is the most important regressor in terms of the magnitude of its e¤ect on migration. In the full model, other important factors in terms of magnitudes are access to electricity and rice price. Income is also important along with language diversity and access to paved roads, but its magnitude is relatively small implying an increase of migration ‡ow by about 0.4 percent in response to a one percent increase in income. When income is allowed to pick up the e¤ect of infrastructure and services, we …nd income to be one of the most important determinant of migration ‡ow next only to distance in terms of magnitude of e¤ect. Even after allowing income to pick up part of the e¤ect of access to electricity, access to electricity still remains as an important determinant of migration ‡ow. Interpreting the second stage coe¢ cient as capturing the amenity value, the elasticity of migration with respect to electricity in the second stage con…rms that migrants do assign considerable value to access to electricity as an amenity. Our …nding regarding access to electricity is consistent with that of Lall, Timmins and Yue (2009). However, unlike Lall, Timmins and Yue (2009) who …nd access to electricity to be valued only by the poorer households, our results suggest that its amenity value does not vary across skill groups of migrants. This is perhaps due to the fact that access to electricity is still limited in Nepal with only a third of wards reporting to have access. In contrast, Brazil has nearly universal geographical coverage for electricity (97 percent of rural areas), and the access issue there is more of …nancial ability to get an electricity connection and paying bills.

Conclusions
In the standard new economic geography models, labor is assumed to be mobile in the medium to longer term (Fujita, Krugman and Venables (1999)). With labor mobility, any regionally targeted policy intervention in these models induces labor movement so as to restore spatial equilibrium. Evaluation of large public investment projects such as transportation, electri…cation and communication on the other hand tends to use spatial variations in outcomes and treatments to estimate returns while ignoring the labor mobility issue. In this paper, we provide evidence on the response of migration to public infrastructure and services using census and household data from a poor developing country, Nepal.
The empirical analysis of this paper incorporates several improvements over the existing literature on the determinants of internal and international migration. The standard model of migration -estimated mostly for international migration -tends to ignore the role of access to public goods and services in the migration decision. Our conceptual model and empirical estimation show that such model tends to overestimate the importance of income in the determination of migration ‡ow due to the positive correlation between income and provision of public goods. Second, while the empirical studies focusing on migrants' destination choice do pay attention to spatial di¤erences in the provision of public goods, they tend to side-step the issue of migrants' non-random selection. There is now a large literature that demonstrates clearly that migrants tend to be di¤erent from non-migrants in terms of both observables and unobservables (Gabriel and Schmitz (1995), Dahl (2002), Akee (2006), Mckenzie, Gibson and Stillman (2010)). Using a nested logit model of utility maximization by the migrants -as suggested by Ortega and Peri (2013) -we derive an empirical speci…cation which corrects for the heterogeneity between migrants and nonmigrants. Third, we make a distinction between the productivity and hence income e¤ect, and amenity value of basic infrastructure such as electri…cation. The income e¤ect arises from its direct e¤ect on …rm and farm productivity whereas the amenity value derives from its use in household activities (e.g. chores/studying/entertainment). Using the correlation between income and access to these public goods, we develop a strategy to provide conservative estimates of their amenity values.
The empirical results show that migrants prefer areas which are nearer to their birth place and have higher income, better access to electricity and paved roads, higher rice and housing prices and greater language diversity. Consistent with the …ndings of Fafchamps and Shilpi (2013), we …nd that when measures of access to basic public goods are added as regressors, the magnitude of income coe¢ cient declines substantially though it still remains statistically signi…cant. This result con…rms that the income coe¢ cient in a standard migration model might be biased upward. We …nd some heterogeneity in the way income, distance and access to a paved road in ‡uence migration for di¤erent skill groups: more skilled migrants are more responsive to income and access to paved road but less responsive to distance relative to unskilled migrants. The results from the two-stage estimation procedure indicate that migrants attach substantial amenity value to access to electricity. Migrants of di¤erent skill levels (primary, secondary and tertiary education level) appear to attach similar amenity values to access to electricity. The results suggest that better access to electricity attracts migrants not only for its positive productivity and income e¤ect but also for its amenity value.
The main …nding of this paper that migrants do respond to access to public goods has important implications for the placement and evaluation of basic public infrastructure and services. While geographical coverage of these public goods should be universal, budget constraints often force governments to prioritize their roll out. Our empirical results suggest that governments can perhaps give more weight to cost considerations in prioritizing the roll out. Our results also suggest that impact evaluation of public investment 23 should pay particular attention to spill-over e¤ects to non-treatment areas due to migration. Such spill-over e¤ects can in turn lead to substantial downward bias in the estimates of returns to public investment when its e¤ect on migration is ignored in the evaluation studies.      Note: All regressions include birth district fixed effects and weighted using destination population. All standard errors are clustered at destination district level. . ELF: Ethno-Linguistic Fractionalization Index. Note: All standard errors are clustered at destination district level.
Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 Note: All regressions include birth district fixed effects and weighted using destination population. All standard errors are clustered at destination district level. Note: All regressions include birth district fixed effects and weighted using destination population. All standard errors are clustered at destination district level. Note: All regressions include birth district fixed effects and weighted using destination population. All standard errors are clustered at destination district level. . ELF: Ethno-Linguistic Fractionalization Index.