Journal of Development Economics Vol. 56 Ž1998. 27–50
Migration, self-selection and earnings in Philippine rural communities Leonardo A. Lanzona
)
Department of Economics, Ateneo de Manila UniÕersity, P.O. Box 154, Manila 1099, Philippines Received 1 April 1996; accepted 1 January 1997
Abstract Estimated returns to schooling investments can be misleading if migration causes significant shifts in population distribution across time. Data gathered in rural Philippine communities show that the more educated and experienced individuals are more likely to outmigrate, causing a sample selection bias in the estimation of wage equations. The observed wages were then lower than the conditional population mean of an entire cohort residing originally in the area. Controlling for self-selection, the wage returns to schooling and experience were higher. Finally, the sample selectivity variable accounts substantially for the difference in the wages of men and women. q 1998 Elsevier Science B.V. All rights reserved. JEL classification: J24; J31; D10 Keywords: Human capital; Migration; Self-selection; Earnings function; Wage gender differentials
1. Introduction Social returns to human capital investments have increasingly been evaluated from analyses of relationships between workers’ earnings and their human capital. However, in communities that experience substantial outmigration, as is usually the case in rural areas in developing countries, the distribution of the population may shift significantly across periods. The wages of those workers who then stay behind in their places of origin may not indicate, without correction for sample )
Corresponding author. Tel.: q632-924-4601; fax: q632-426-5601
0304-3878r98r$19.00 q 1998 Elsevier Science B.V. All rights reserved. PII S 0 3 0 4 - 3 8 7 8 Ž 9 8 . 0 0 0 5 1 - 0
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
28
selectivity bias, the true payoff to human capital investments of an entire cohort that originally resided in these areas ŽSchultz, 1988; Behrman and Birdsall, 1988.. Hence, the earnings observed among nonmigrants may not be random samples of the original populations, but rather censored samples. The failure to consider migration endogenously in the wage estimates can result in biased estimates that can lead to misleading evaluations of government programs ŽRosenzweig and Wolpin, 1986.. If returns to public schooling investments and post-schooling training programs are evaluated to be low in the villages experiencing substantial outmigration, such investments may be diverted to other areas. Hence, calculations of returns to schooling from unconditional cohort wage estimates will have important implications for the distribution of local human capital investments. Moreover, if these migrants move towards urban areas, the observed outmigration of more educated and potentially more experienced individuals will cause rural areas to remain poorer than the destination areas, unless remittances are large. The society in general is not made worse off since the more productive workers are given higher earnings in the urban areas. However, if the urban areas retain their educated children and, consequently, their earnings while the rural areas lose many of them, after having locally financed them, the costs of public schooling investments for each rural household can be greater than those for urban households. Consequently, the intergenerational income distribution across regions will be less equitable with the rural areas carrying heavier tax burdens. The problem of sample selection was first recognized in estimates of female wages because of women’s limited participation in the wage labor market. In agricultural and low-income communities, access to land and imperfections in the wage labor market may also reduce the wage labor market participation of both men and women Žsee, for example, Chiswick, 1976; Anderson, 1982.. Self-selection thus appears to be more pronounced in less-developed countries where self-employment is common. Furthermore, in studies using longitudinal household data, another source of selection bias is children’s decisions to marry and leave the parental home. Only those who remain in their original households are actually resurveyed, possibly resulting in a further nonrandom sample of the wage earners in the villages. Nevertheless, the individual’s decision to leave home—whether due to marriage or to job migration—is caused by his increased earnings capacity in addition to the Žnet. consumption costs of remaining in an extended family. 1 Hence, the 1
While several studies view marriage as a strong predictor of leaving the parental home, individuals generally do not marry unless they have the ability to establish their own household. In this case, marriage and the decision to leave the parental home can be simultaneously determined and, therefore, conditioning on marital status is not a solution Žsee Becker, 1981.. Furthermore, in societies where extended families are commonplace, the correlation between marriage and the decision to leave home may be low.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
29
attrition of a birth cohort due to household members’ decision to leave the parental home can just be a form of self-selection through migration. The key issue here is that these sources of wage attrition may be related to other unobserved factors, such as individual ability, that affect earnings. In particular, migration is a human capital investment process that may be endogenous to the earnings function. If so, such investments must be included in the earnings function, but must be explained by instrumental variables that are uncorrelated with ability, preferences or other unobservables. The paper presents an analysis of the behavior of migrants and non-migrants in rural low-income economies in order to determine the effect of migration selectivity on wages. 2 One way of doing this is by incorporating a selection control variable into wage estimates of persons who reported wages in the villages ŽHeckman, 1979.. Since a number of job and locational choices are available to a worker, the paper uses a multinomial logit–ordinary least squares ŽOLS. two-stage estimation of wages ŽLee, 1983.. A particular locational choice considered in the model is the case in which the individual leaves his parental residence, but remains in the same village. Models that consider migration under uncertainty Že.g., Harris and Sabot, 1982. posit workers as looking for jobs in a given locality without complete knowledge of the characteristics of that locality. The features of the locality can only be known by living there. These models can be extended by assuming that the migrant is choosing among alternative residential options once he has chosen to accept a job in a larger given locality. It can be hypothesized that, given the skills they possess, the more educated and abler individuals are more likely to move to areas farther from their parental homes because their expected wages are higher, and the more specialized benefits they can obtain will be greater than the costs and risks of long-distance migration. To distinguish short-distance from long-distance migration, differences in personal characteristics and their effects on migration are considered. This framework will be tested on the wage data recently collected in the Philippines. Of particular interest in this data are explanatory variables that had been gathered from previous survey rounds on the same individuals. In many studies, migration that has occurred at different and long periods in the past is associated with independent variables taken at a particular point in time. If these
2 Several previous papers on migration have found positive selection in the wage functions of migrants and nonmigrants Že.g., Nakosteen and Zimmer, 1980; Falaris, 1987.. These papers however do not deal with all of the above sources of wage attrition. Since self-employment and movements within the village are choice variables, treating them exogenously or restricting the sample to wage earners may yield biased results.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
30
regressors change over time, this results in misspecification. 3 To a certain extent, the Philippine data allow for the analysis of migration and work decisions at the time when these decisions are being made. The rest of the paper is organized as follows. Section 2 presents an econometric framework of wage determination where the wage labor participation decision and migration are jointly considered in correcting for sample selection. Section 3 discusses the empirical strategy and variables used to implement the framework. Section 4 describes the data, and Section 5 features the results of the empirical analysis. Section 6 provides concluding remarks and policy implications of the findings.
2. An empirical framework for estimating wages In standard migration theories, individuals are assumed to maximize incomes as producers and sort themselves into more productive jobs ŽHarris and Todaro, 1970; Fields, 1975.. Extensions to this theory consider migration as a mechanism for matching in consumer heterogeneity with regional variations in prices and wages and access to public programs ŽRosenzweig and Wolpin, 1988; Schultz, 1988.. From these factors, a wage framework can be formulated by the following equations: ln wsi s Z sihs q m si Ysi ) s x si b s y d si g q e si
Ž 1.
where Z si refers to the vector of exogenous variables that determine Žlogarithmic. wages Žln wsi ., and x si and d si are distinct vectors of independent variables that affect the probability of choosing sth option Ž Ysi . which, in this case, is participation in the local wage labor market. The vector x si stands for observed individual characteristics Že.g., age and schooling. that pertain to the person i’s productivity. The vector d si refers to distance of i’s household from urban center, such as schools and hospitals, that reflects the costs and benefits of locating in a particular area or household ŽSchwartz, 1973.. We assume that EŽ m si < Z si , x si , d si . s 0. Note that wages can be observed only if the individual is a non-migrant and a wage earner. Assume that there are J number of job choices Ž j s 1, . . . , J .. In the data, the individual has four choices: Ža. work in the local labor market; Žb. be self-em3
Greenwood Ž1971. estimated a migration model for India, employing both lifetime and one-year migration flows as dependent variables and relating them to the same set of exogenous variables. Absolute values of his regression coefficients were consistently higher for the lifetime form of the dependent variable.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
31
ployed; Žc. migrate from the parental home to work in some other location or Žd. somewhere in the same village. 4 A basic premise here is that the same forces that determine the choice of workers to migrate from rural to urban areas can also explain their choice between one labor market and another within the rural or the urban sector. The individual decision maker is presumed to consider all labor market opportunities and locational amenities and to choose one that maximizes his utility. Given these choices, Yji is a polychotomous variable with discrete values. Yji is equal to Ysi ) given the following condition: Yji s Ysi ) iff Ž x si b s y x ji b j . q Ž d ji g j y d si gs . ) e si y e ji for every choice j. This means that Yji s Ysi ) if and only if Ysi ) ) Max Yji . Ordinary least squares ŽOLS. estimates of wage equations can then be biased if a covariance between the error terms of the labor market participation decision and the wage functions exists ŽHeckman, 1979; Maddala, 1983.. If the probability of being in the wage labor market is a significant factor in explaining the observed wages, then the OLS-estimated wage equation will conditional only on the characteristics of those who reported wages. The paper assumes that wage labor market participation is determined independently of the decision to migrate from the parental home village or to engage in self-employment in the same parental home. If the choices are neither substitutes nor complements, or are mutually exclusive, then the probability of being in the wage labor market can be estimated consistently using the multinomial logit ŽMNL. approach suggested by Lee Ž1983.. This model exhibits the property of independence of irrelevant alternatives ŽIIA., suggesting that the changes in opportunities in one option will not influence the other options. 5 While various factors influence the choices of households in the rural economy, the job choices themselves may often be distinct and homogeneous, especially in a low-income economy where complex heterogeneous and interrelated choices may not be found. Hence, there is no a priori reason for ruling out the assumption of IIA. Nevertheless, since this may be a restrictive property, a test of the validity of this assumption will be performed. 4
Given insufficient data on the earnings of those who left their residences, individuals who migrate from their parental home are assumed to engage in productive work. Given the various seasons and production activities in the rural economy, work and hence wages in farms even in one village may also be presumed to vary. 5 More precisely, when the probability of one option changes, the probabilities of the other options change proportionately so that the ratio of their probabilities remain constant, leaving the relative importance of the choices unaffected. This model has been used to analyze situations dealing with multiple economic choices within the framework of utility maximization Žsee McFadden, 1974.. Furthermore, this model is of practical use if the objective is simply to examine choices among a subset of alternatives rather than among all alternatives open to the worker ŽTrain, 1993..
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
32
Furthermore, since the multinomial logit procedure for estimating labor participation requires that the error term, e si , be independently and identically distributed with the Weibull distribution, the Lee Ž1983. method performs a change of variable on the choice probabilities from a logistic distribution to a standard normal distribution. One way of characterizing the Lee model is the following: ln wi s Z ih q m i when Ii ) s 1 Ii ) s
½
1 if x i u y d i v q e i ) ) 0 0 otherwise
Ž 2.
where
mi e i ) ; NID
ž
2 0 sm , 0 sme
1
/
and h , u and v refer to the estimated parameters, and e i ) is a transformed standard normal error term with zero mean and unit variance. Let PrŽ Ii ) s 1. s PrŽ Ji s s. where
°1 if V
1 i q e 1 i ) Vsi q e si for all s / 1 2 if V2 i q e 2 i ) Vsi q e si for all s / 2 Ji s ...
~
¢J if V q e ) V q e ji
ji
si
si
for all s / J
e ji , j s 1, . . . , J are independently and identically distributed with the Weibull distribution, and Vji are parametric functions of the observed characteristics and distances. Lee Ž1983. showed that, conditional on the alternative s being chosen, i.e., work in the local labor market, we can estimate wages based on the following: ln wsi
s
Z sihs q dem lŽ csi . q m si
Hsi
s
H Ž csi . s Gy1 Ž csi .
lŽ csi .
s
g Ž Hsi .
Ž 3.
G Ž Hsi .
Hsi is the inverse of the standard normal distribution evaluated at w csi s ProbŽ Ji s s .x and the functions g Ž.. and GŽ.. are the probability density and the cumulative density functions of the standard normal distribution. The term csi is equivalent to the probability of wage market participation or the possibility of observing a wage, and is estimated by multinomial logit. The term dem refers to the product of the standard deviation of m si and the covariance
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
33
between e i ) and m si . This selectivity control term, l si , is significant if unobserved worker ability or heterogeneity exist, and if the workers’ wages are affected by such abilities and heterogeneity. In Eq. Ž3., the covariance term in sem is unrestricted. A negative covariance indicates a negative selection bias, suggesting that those who decide to stay in the parental home and engage in wage jobs receive lower wages than the population mean. A positive covariance supports the comparative advantage hypothesis that individuals choose the option that optimizes their productivity. Estimating wage functions that omit the selectivity control term may result in biased measurements of the marginal effects of regressors in the wage function. The bias, if any, is determined by the influence of this particular independent variable on l si . The full effect of a regressor, z, on wages can be written as ŽGreene, 1993; Maddala, 1983.:
E E Ž ln wsi < Ji s s . Ez
s hz y b z sem r Ž csi .
Ž 4.
where r Ž csi . is the effect of a change in csi on l si , hz is the direct effect of z on wsi , and b z is the effect of effect of z on labor market participation. Since r is positive if labor market participation has occurred, the sign of the expression after the minus sign in Eq. Ž4. depends on the signs of b z and sem . If there is a negative selection bias Ži.e., sem - 0. and if b z is also negative, then the effect of the regressor z will be underestimated in an estimation based on the given sample. In the case of education, if sem and b z are negative, then the estimated coefficient of education is lower than would be the case if the selectivity control variable were correctly included.
3. Empirical strategy and econometric model This study uses the Bicol Multipurpose Surveys ŽBMS. conducted in 1978, 1983 and 1994. The surveys cover several economic and geographic units in the Bicol Region of the Philippines and contain information on household assets, individual characteristics and community variables. The analysis, however, will focus on the 59 villages composed of 691 households that were included in all three surveys. Of the sample village clusters, 40 are located in rural areas, 9 are in towns Žor poblaciones. and 10 are in cities. In most studies, the dependent variable used to measure individual migration runs for long periods Ži.e., lifetime. and is related to the explanatory variables defined at a particular period of time. For instance, estimates using census data usually measure migration behavior that has occurred over several periods. The problem is perhaps even more serious in household surveys. Individual migration is usually taken from different, separate periods in the past, but the regressors may
34
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
be taken only from the time of the survey. This approach leads to a misspecification if the explanatory variables change across time. Furthermore, because the explanatory variables may be consequences of past migration and work decisions, a simultaneous equation bias results from the misspecification. Several studies suggest these models will usually overstate the effects of the explanatory variables, relative to models that employ current migration and work decisions Žsee Levy and Wadycki, 1974.. The procedure then is to consider migration decisions at periods when individuals are prone to migrate, and when locational and work decisions are not able to determine the explanatory variables. A way to deal with this misspecification is to use lifetime mobility and occupation histories of individuals. However, given only three survey periods, an alternative approach is to pool data from the three surveys so that the explanatory variables used for the individual observation will be taken from the survey period when the person was at risk of migrating. Moreover, since parents have already decided on their work and location at the time of survey, and since they have influenced the independent variables to some extent, they will be excluded from the sample. Thus, the empirical analysis will include only children, age 15 to 40 years, who had completed their schooling at the time of the survey. Specifically, the empirical analysis in this paper will include children who were between 15 and 40 years old in 1978, 15 and 20 years old in 1983, and 15 and 25 years old in 1994. Differences in age ranges are due to the time span between surveys and the objective of not including any child twice. Fig. 1 shows a standard cohort table and illustrates the sampling procedure used. Note that the diagonal age ranges shown refer to the same individuals. In effect, individuals who are 26 or older in 1983 are not considered because they were already part of the 1978 cohort. Hence, a pooled sample of 2905 persons, aged 15–56 in 1994, can be analyzed using variables taken from previous surveys. This approach assumes that the locational and occupation choices made at these particular periods are permanent. Although this may not be restrictive for migra-
Fig. 1. Sampling procedure.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
35
tion decisions, it may be a source of bias for estimates of wage labor market participation if individuals’ engagement in these markets varies. Because of this, the analysis will only consider the wage data for people who have been employed permanently in the labor market on a full-time basis. In the case of the younger cohort, only those who have been usually employed or worked on longer term contracts will be considered. These restrictions on the data imply that since the sample used may no longer be random, the results will no longer hold for everyone in the villages. In any case, the approach provides a consistent procedure for examining the effects of migration on earnings. Another task of this paper is to identify variables that are distinct for wage labor market participation and wages. The maximum likelihood estimates of Eq. Ž1. are identifiable if there is at least one independent variable in the choice functions but not in the wage equation ŽMaddala, 1983.. Factors that affect the reservation wages for nonwage earners and migrants but do not affect the market wages would then identify the model. These factors include the household’s nonearned income, land and other assets that influence the persons’ productivity in self-employment or the individuals’ demand for leisure, given leisure is a normal good. In addition, distances from households to public services, specifically schools, are to be included in order to identify further the migration decisions. 6 Based on the empirical framework, the reduced-form wage labor participation probability and Žlogarithmic. wage equations to be estimated for both men and women are as follows: Prob Ž Y s wage earner . s f Ž Human Capital, ParentsX Education, Household Assets, Distance, Cohort. log Ž Wage . s g Ž Experience, Experience Squared, Schooling.
Ž 5.
where human capital refers to schooling and potential experience indicated by age; parent’s education refers to the completed schooling of the mother and father; household assets include nonearned income, market value of parental home and land; distance is the distance from the household to public services, specifically schooling; and cohort refers to the year of the survey in which the observation was collected. Given the available data, the following variables will be used. 3.1. Human capital For both labor participation and migration, higher education is expected to result in higher rewards for both worker and migrant ŽBecker, 1993; Sjaastad, 6 The model presented here is actually overidentified which is somewhat desirable because it ensures good finite-sample properties for the selectivity control variable. Moreover, the validity of choice of variables in an overidentified model may be tested to some extent, as will be done later.
36
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
1962; Schultz, 1982.. Education here is defined as the number of completed schooling years and is assumed to capture the skills the individual may bring to a given job in the outside market. Previous studies have also shown that migration is influenced inversely by a person’s age. Older people are less likely to migrate since they have less time to pay back investments. To control for these effects, dummy variables are included for persons between 15 and 19 years of age, and for persons between 20 and 24. In the estimates for all children, a dummy variable for women is included. Women who have may more to gain from leaving the rural areas are expected to differ from men in their propensity to migrate. 3.2. Parents’ education The years of schooling of parents are assumed to influence children’s decisions in two ways. First, these variables capture unobserved family background effects that can affect the choice of whether the child stays in the house or migrates. For instance, households with more educated parents can possess more information about a particular locality and so induce greater migration ŽMincer, 1978.. Second, these variables may correlated with various assets, including family connections, that can lead to greater self-employment activities or leisure. 3.3. Household assets If household production and consumption decisions are separable, the returns from household production activities can be considered within an agricultural household framework where markets are available ŽSingh et al., 1986.. In this case, prices can equally affect both those who are engaged in the labor market and those who are not. Hence, the returns from self-employment would be solely derived from household resources. Studies Že.g., Evenson and Binswanger, 1984. have shown that agricultural household resources lead to greater demand for household labor. The effects of ownership of productive assets, for example, land and the amount of irrigated land, on wage labor participation and migration are expected to be negative as demand for household labor increases. Labor participation and migration can be determined by the market value of the house and the household’s unearned income measured by a nonland earning asset dummy. These variables indicate the household’s relative wealth which positively affects higher reservation wages and may lead to lower participation in the labor market. 3.4. Distances and origin Household distances to schooling can capture the non-wage differential costs and benefits from migration. The variables included are the distance from the house to primary and secondary schools Žin kilometers.. If the benefits of moving
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
37
closer to these public services are greater than the costs, individuals are likely to migrate. The origin of the individual may also indicate an individual’s propensity to migrate. People from rural areas are more likely to migrate in order to obtain higher wages and to acquire goods and services found mostly in urban areas. 3.5. Cohort The survey years from which the data have been taken will be included in the logit functions in order to account for the so-called cohort effects. It is expected that individuals born at a particular period of time face the same market and social conditions relative to those born in other periods. Dummy variables were assigned to the 1978 and 1983 cohorts. For the wage equation, the dependent variables are the wage rates reported in the 1994 survey. The independent variables are the standard Mincerian wage function variables, i.e., potential labor market experience and years of schooling for education Žsee Willis, 1986, for a survey.. Potential experience is measured by getting the difference between the worker’s working age and the age of school completion.
4. Descriptive statistics The survey site is Camarines Sur, the main province of Bicol, which is considered one of the poorest regions in the country. In 1994, while economic conditions had somewhat improved, Bicol had the lowest per capita Gross Domestic Product, the highest poverty rate and one of the lowest in-migration rates in the country. Its terrain is mostly mountainous, with a limited amount of irrigation. A number of communities are inaccessible and isolated. While elementary schools are generally found in each village, they tend to be located close to the village center. High schools are usually located in some village towns and larger urban areas. To a certain extent, the degree of urbanization is manifested by the presence of higher educational institutions. Table 1 features the total percentages of children who had left the household as of the 1994 survey period by each age, sex and schooling cohort. The table is then subdivided into those who remained in the village and those who migrated. The following points can be noted from Table 1. First, daughters appear to lead sons in moving out of the household. This seems to be particularly significant in the case of outmigration. Second, education has a perceptible effect on the decision to leave. Years of schooling is negatively related to the percent staying in the village, but is positively related to the rate of moving out, especially for sons. The age groupings in the table identify the different birth cohorts in the analysis. The table shows that 84% of those born between 1938 and 1963 Žthe
38
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
Table 1 Percentage of children leaving parental home as of 1994 Žin percentage of age, sex and education cohort, 1978–1994 Bicol surveys. Years of schooling
Age categories 21–25
26–30
Over 30
Total
I. Total no. of children leaÕing (in percent) Women 0–6 40.32 7–10 47.47 Above 10 9.09 Sub-total 42.44 Men 0–6 20.69 7–10 21.62 Above 10 25.00 Sub-total 21.21 Total 31.08
15–20
67.68 62.10 44.07 60.28 55.63 51.46 53.19 53.77 56.97
83.15 79.66 72.22 80.72 69.72 66.00 66.67 68.42 74.48
88.75 84.54 81.22 85.85 85.58 82.92 73.91 82.57 84.15
80.27 70.76 70.18 75.02 68.63 65.94 67.54 67.58 71.17
II. Stayed in the Õillage Women 0–6 7–10 Above 10 Sub-total Men 0–6 7–10 Above 10 Sub-total Total
3.23 2.02 0.00 2.33 2.59 0.00 0.00 1.52 1.89
21.21 12.90 6.78 14.54 12.68 10.68 6.38 10.96 12.72
26.97 18.64 5.56 21.69 26.61 26.00 0.00 24.56 23.15
33.50 24.64 13.71 26.45 37.21 25.62 11.80 28.78 27.66
27.92 16.36 11.23 20.66 26.35 18.90 9.65 21.40 21.04
III. Outmigrants Women 0–6 7–10 Above 10 Sub-total Men 0–6 7–10 Above 10 Sub-total Total
37.10 45.45 9.09 40.12 18.10 21.62 25.00 19.70 29.19
46.46 49.19 37.29 45.74 42.96 40.78 46.81 42.81 44.25
56.18 61.02 66.67 59.04 43.12 40.00 66.67 43.86 51.34
55.26 59.90 67.51 59.41 48.37 57.30 62.11 53.78 56.50
52.35 54.40 58.95 54.36 42.28 47.05 57.89 46.18 50.13
The figures show the percentage for each sex, schooling and age cohort. For each cohort, the denominator is the total number of children, aged 15 and above, who are still living and have already completed schooling.
1978 birth cohort. have left their parents’ home, with 57% moving out of the village. Seventy-four percent of the children born between 1964 and 1968 Žthe 1983 birth cohort. have also left their parental homes; 51% of them have moved out to another village. Younger persons Žbelonging to the 1994 cohort. seem to have a relatively lower rate of moving out. Only 57% of those born between 1969 and 1973 have left home, and 44% of this cohort migrated out of the village. Only
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
39
29% of the youngest cohort Žthose born between 1974 and 1979. have left for other villages. To some extent, these patterns capture the household’s maturation as children become older and establish their own households. However, economic developments in the country also suggest that conditions are better in recent years than in the previous periods. Table 2 provides the labor force participation rate of those remaining in the village for each age, sex and schooling cohort. Three main points can be noted. First, the effect of higher education does not seem prominent, especially for the men. Roughly 45% of the male workers have only completed high school. For the women, higher education is associated with greater rates of labor market participation, but the difference in the rates across these educational levels is not substantial. Second, individuals belonging to the 21 to 25 and 26 to 30 age groups appear more likely to engage in wage labor. Third, women participated less than men in the wage labor market. The means and standard deviations of the variables for the male and female subpopulations and for the whole sample for the three survey years are shown in Table 3. As has already been inferred from Table 1, migration is highest for the 1978 cohort followed by the 1983 cohort. The younger 1994 cohort however has the highest percentage of wage job participation for the entire sample. Wages in 1994 are recorded to be highest for the 1978 cohort with the 1983 cohort having the lowest mean wage in the data. For the independent variables, the following findings are noteworthy. First, the average years of schooling completed are highest for the 1994 cohort, followed by the 1978 cohort. Second, the 1994 cohort appears to come from households with more assets. A greater proportion of them have nonland earning assets, and the discounted market value of their houses, based on the Consumer Price Index
Table 2 Labor market participation rate by age, schooling and sex Žin percent of age, sex and education cohort, 1994 Bicol Survey.
Women
Men
Years of schooling
Age in 1994 15–20
21–25
26–30
Over 30
0–6 7–10 Above 10 Sub-total 0–6 7–10 Above 10 Sub-total
24.32 32.69 20.00 28.28 43.48 41.38 0.00 41.03 36.08
37.50 42.55 42.42 41.07 38.10 56.00 40.91 45.19 43.32
20.00 25.00 60.00 28.13 51.52 29.41 50.00 44.44 38.37
19.57 9.38 40.54 23.48 29.03 41.67 33.33 34.21 29.59
Total See notes in Table 1.
Total
25.38 30.07 40.00 30.73 39.60 44.51 33.78 40.44 36.37
40
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
Table 3 Means and standard deviations of variables Variable description Cohorts 1978 All
1983 All
Women Men
All
Women Men
Dependent Õariables Proportion of children leaving parental home Outside of village 0.554 0.583 0.527 Ž0.50. Ž0.49. Ž0.50. Within village 0.296 0.281 0.309 Ž0.46. Ž0.45. Ž0.46. Proportion reporting 0.042 0.034 0.049 Ž0.20. Ž0.18. Ž0.22. wage job Reported daily wage 113.02 129.05 103.22 Ž84.45. Ž100.06. Ž73.13. in 1994 Žin pesos.
0.510 Ž0.50. 0.245 Ž0.43. 0.087 Ž0.28. 76.88 Ž48.10.
0.595 Ž0.49. 0.223 Ž0.42. 0.041 Ž0.20. 73.33 Ž54.16.
0.427 Ž0.50. 0.267 Ž0.44. 0.133 Ž0.34. 78.21 Ž46.82.
0.431 Ž0.50. 0.108 Ž0.31. 0.189 Ž0.39. 83.01 Ž68.75.
0.481 Ž0.50. 0.125 Ž0.33. 0.146 Ž0.35. 74.57 Ž71.71.
0.381 Ž0.49. 0.094 Ž0.29. 0.229 Ž0.42. 88.18 Ž66.62.
Human capital Age dummy Žduring survey. 15–19 0.231 Ž0.42. 20–24 0.273 Ž0.45. Years of schooling 7.767 Ž3.26.
0.235 Ž0.42. 0.267 Ž0.44. 7.759 Ž3.35.
0.227 Ž0.42. 0.280 Ž0.45. 7.776 Ž3.18.
1.000 Ž0.00. 0.000 Ž0.00. 7.151 Ž2.66.
1.000 Ž0.00. 0.000 Ž0.00. 7.351 Ž2.75.
1.000 Ž0.00. 0.000 Ž0.00. 6.953 Ž2.56.
0.230 Ž0.42. 0.424 Ž0.49. 8.492 Ž3.34.
0.223 Ž0.42. 0.429 Ž0.50. 9.048 Ž3.31.
0.235 Ž0.42. 0.421 Ž0.49. 7.962 Ž3.28.
Parents’ education Mother’s education Žyears. Fathers education Žyears.
4.350 Ž2.92. 4.501 Ž3.16.
4.360 Ž2.80. 4.559 Ž3.13.
4.339 Ž3.02. 4.445 Ž3.19.
4.966 Ž2.33. 4.785 Ž2.53.
4.845 Ž2.46. 4.703 Ž2.43.
5.087 Ž2.20. 4.867 Ž2.63.
5.811 Ž2.77. 5.867 Ž2.98.
5.954 Ž2.86. 5.936 Ž2.94.
5.678 Ž2.67. 5.807 Ž3.02.
0.101 Ž0.30. 0.033 Ž0.08. 1.859 Ž4.74. 0.189 Ž0.76.
0.097 Ž0.30. 0.033 Ž0.08. 1.561 Ž4.04. 0.206 Ž0.82.
0.105 Ž0.31. 0.034 Ž0.08. 2.127 Ž5.28. 0.173 Ž0.70.
0.044 Ž0.20. 0.018 Ž0.04. 0.717 Ž2.44. 0.072 Ž0.51.
0.041 Ž0.20. 0.019 Ž0.05. 0.680 Ž2.59. 0.047 Ž0.30.
0.047 Ž0.21. 0.016 Ž0.03. 0.754 Ž2.30. 0.097 Ž0.65.
0.138 Ž0.35. 0.061 Ž0.15. 1.404 Ž3.62. 0.306 Ž0.85.
0.148 Ž0.36. 0.064 Ž0.15. 1.279 Ž3.47. 0.283 Ž0.80.
0.129 Ž0.34. 0.058 Ž0.15. 1.522 Ž3.76. 0.328 Ž0.88.
0.737 Ž0.91. 3.305 Ž4.02.
0.786 Ž0.92. 3.327 Ž4.30.
0.695 Ž0.89. 3.302 Ž3.78.
0.795 Ž0.95. 3.485 Ž3.98.
0.806 Ž0.81. 3.791 Ž4.11.
0.783 Ž1.08. 3.183 Ž3.85.
0.658 Ž0.94. 2.410 Ž3.44.
0.635 Ž1.00. 2.438 Ž3.97.
0.681 Ž0.88. 2.382 Ž2.85.
0.731 Ž0.44. 1362
0.724 Ž0.45. 648
0.737 Ž0.44. 715
0.815 Ž0.39. 298
0.818 Ž0.39. 148
0.813 Ž0.39. 150
0.742 Ž0.44. 1245
0.719 Ž0.45. 609
0.765 Ž0.42. 637
Household assets Nonland earning asset dummy Value of house Žpesos.r100,000 Land area Žha. Irrigated land owned Žha. Distance and origin Distance from primary school Žkm. Distance from secondary school Žkm. Origin, rurals1 No. of observations
Women Men
1994
Figures in parentheses are standard deviations.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
41
ŽCPI., is roughly twice that of 1978. These persons also belong in households with more irrigated land areas. Third, the 1994 cohort appears to reside closest to primary and secondary schools; the 1983 cohort lives farthest from the schools. The 1983 cohort also has the highest percentage of individuals who were originally from rural areas. These results suggest that more educated and relatively experienced individuals who are about to start their own working careers and families tend to migrate outside of the village. Interestingly, the women appear to have taken the lead in outmigration. Moreover, as expected, these migrants, on average, have resided in households which own limited assets, i.e., in terms of irrigated land and house values, and which are located in areas farthest from schools.
5. Results of the econometric analysis Table’s 4A and B present the regression coefficients and t-values from the MNL model for males and females. Tests based on a chi-square statistic proposed by Hausman and McFadden Ž1984. failed to reject the null hypothesis that independence of irrelevant alternative holds at greater than 99.9% level. The test involves estimating separate binary logit models from the household’s choice set, using the choice of staying at home and not being a wage earner as a reference. The estimates from the binary logit models were found similar to the coefficients of the full multinomial logit model, suggesting that no complementarity or substitution is found for each of the choices. The coefficients show the effects of the regressors on the probability of participating in the wage labor market, leaving the parental home and staying in the village, or migrating to another location, relative to the likelihood of staying at home and not being a wage earner. Confirming the findings in the first three tables, the most noteworthy of these results are the following. For the human capital variables, age and education have significant effects on the decision to leave the parental home. Individuals have an increasing propensity to leave their parental home as they reach the age of 25. Women migrate outside of the village at a later time than men. Completed schooling years also show a significant positive effect on the decision to leave the village, especially for the women. Moreover, as expected, less educated persons who leave their parents’ house show a greater propensity to stay in the village. In the case of the women, education also raises the probability of engaging in wage activities. The more educated women are more likely to either leave the village or stay in their parental home and become wage earners. For men, the presence of nonland earning assets as well as higher values of the house induce a negative effect on the likelihood of moving out or engaging in a wage job. This can be due to greater self-employment activities in the parental
42
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
home as result of greater household assets. For females, the value of the house and ownership of irrigated land are also negatively associated with the decision to leave the village because perhaps of its positive effect on self-employment.
Table 4 A. Multinomial logit model for males Variables
Wage earner
Left parental home Stayed in the village
Constant Age dummy Žduring survey. 15–19
y0.3233 Ž0.72.
0.2817 Ž1.04. 20–24 0.4057 Ž1.57. Years of schooling 0.0081 Ž0.25. Mother’s education y0.0180 Ž0.42. Father’s education 0.0375 Ž0.95. Nonland earning asset dummy y0.6636)) Ž2.07. Value of house Žat 1978 prices. y0.8045 Ž1.14. Area of land owned Žha. y0.0116 Ž0.44. Irrigated land Žha. y0.1080 Ž0.84. Distance from primary schools Žkm. 0.1491 Ž1.22. Distance from secondary schools Žkm. y0.0809)) Ž1.99. Origin, rurals1 y0.2377 Ž1.01. Year data collected 1978 y0.3767 Ž1.54. 1983 0.0091 Ž0.03. Log-likelihood y1696.25 Restricted Žslopess 0. log-likelihood y1912.19 Chi-squared Ž42. 431.88 Significance level 0.00 N 1502
y0.4746 Ž1.12.
Outside the village 0.1311 Ž0.37.
y1.6004)) Ž6.40. y0.7061)) Ž3.09. y0.0724)) Ž2.34. y0.0117 Ž0.30. 0.0329 Ž0.90. y0.6035) Ž1.92. y3.3933)) Ž2.33. y0.0216 Ž0.97. 0.0928 Ž0.92. 0.2461)) Ž2.23. y0.0265 Ž0.94. 0.6079)) Ž2.53.
y1.3513)) Ž6.39. y0.2921 Ž1.54. 0.0482) Ž1.90. y0.0324 Ž0.97. 0.0367 Ž1.19. y0.5409)) Ž2.27. y1.0303) Ž1.71. y0.0058 Ž0.35. y0.1333 Ž1.46. 0.2451)) Ž2.50. 0.0043 Ž0.18. 0.1176 Ž0.61.
2.3264)) Ž10.75. 2.3868)) Ž6.87.
1.4848)) Ž8.57. 1.5535)) Ž5.29.
Figures in parentheses are absolute values of asymptotic t-values. )), ) refer to 5% and 10% levels of significance, respectively.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
43
Table 4 Žcontinued. B. Multinomial logit model for females Variables
Wage earner
Left parental home Stayed in the village
Constant Age dummy Žduring survey. 15–19
y1.4998)) Ž2.60.
0.1316 Ž0.39. 20–24 0.4358 Ž1.42. Years of schooling 0.0770) Ž1.88. Mother’s education y0.0048 Ž0.09. Father’s education 0.0397 Ž0.77. Nonland earning asset dummy 0.2846 Ž0.75. Value of house Žat 1978 prices. y2.5881)) Ž2.13. Area of land owned Žha. 0.0153 Ž0.44. Irrigated land Žha. y0.0599 Ž0.43. Distance from primary schools Žkm. y0.0005 Ž0.00. Distance from secondary schools Žkm. 0.0036 Ž0.10. Origin, rurals1 y0.3358 Ž1.20. Year data collected 1978 y0.3985 Ž1.32. 1983 0.3382 Ž0.64. Log-likelihood y1513.76 Restricted Žslopess 0. Log-likelihood y1639.98 Chi-squared Ž42. 252.43 Significance level 0.00 N 1405
y0.5080 Ž1.13.
Outside the village 0.2827 Ž0.77.
y1.3562)) Ž5.08. y0.2530 Ž1.11. y0.0016 Ž0.05. y0.0179 Ž0.44. y0.0093 Ž0.24. 0.4851 Ž1.50. y4.0955)) Ž2.84. 0.0173 Ž0.62. y0.0561 Ž0.52. 0.0678 Ž0.61. 0.0088 Ž0.35. 0.4513) Ž1.88.
y0.7053)) Ž3.38. y0.3531) Ž1.79. 0.0631)) Ž2.43. y0.0471 Ž1.37. 0.0454 Ž1.37. 0.2311 Ž0.86. y0.9068 Ž1.58. 0.0055 Ž0.23. y0.2348)) Ž2.48. 0.1470) Ž1.64. 0.0199 Ž0.91. y0.0324 Ž0.17.
1.6719)) Ž7.78. 2.0304)) Ž5.44.
1.1561)) Ž6.55. 1.2647)) Ž4.27.
See notes in A.
The regression results show that greater distances from key urban centers, such as primary schools, induce more migration, indicating that the benefits of nearby schools reduce the net gains to migration. Especially for the men, access to primary schools appears to be a goal of migration. The likelihood of becoming a
44
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
wage earner in areas away from secondary schools is also lower for men. Having origins in a rural area seems to increase the probability of moving out of the parental home. The variables showing cohort effects indicate that those who were sampled from the 1978 and the 1983 surveys have a greater tendency to move out of their parental home and leave the village. This can perhaps be attributed to the recent improved economic conditions in the area. Table 5 reports both the ordinary least square ŽOLS., and the sample selectioncorrected wage estimates. Differences can be observed for the estimated returns in the independent variables as well as in statistical significance as shown by the t-values. There is significant negative selection bias for both men and women, or in other words, the wages of persons staying in the parental home and engaging in wage jobs are lower than expected on the basis of the observables. This supports the hypothesis that workers who remain in the village are less productive than those who migrate. However, the exact reason for lower productivity cannot be determined by the model since ability, motivation, and wealth are all subsumed in the selectivity control term. An equally important finding in Table 5 is the underestimation by OLS of the returns to experience for all individuals and to schooling in the case of men. With the selection-corrected estimates, slightly greater returns to experience are seen for all workers. In the case of education, returns are higher for the men, and about the
Table 5 Wage estimates
Constant Experience Experience squaredr100 Years of schooling Lambda Žselection control. Adjusted R 2 F-test Wald test N
Ordinary least squares
Selection-controlled estimates
Males
Females
Males
Females
3.3702)) Ž21.32. 0.0474)) Ž2.90. y0.1013)) Ž2.03. 0.0752)) Ž5.45. y
2.6585)) Ž11.19. 0.0672)) Ž3.02. y0.0963 Ž1.54. 0.1026)) Ž5.40. y
3.7202)) Ž21.37. 0.0686)) Ž5.65. y0.1103)) Ž7.22. 0.0893)) Ž11.33. y0.4682)) Ž7.34. 0.17 10.59 123.68 187
3.8138)) Ž17.22. 0.0913)) Ž10.23. y0.1018)) Ž8.65. 0.0994)) Ž10.90. y0.7895)) Ž7.53. 0.33 13.20
0.16 12.84 39.87 187
0.30 15.22 102
102
Figures in parentheses are absolute values of asymptotic t-values. )),) refer to 5 and 10% levels of significance, respectively. The Wald test is used to determine the structural difference between the estimates for males and females.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
45
same for women. This result is expected since the relatively experienced and more educated individuals have left the villages. If more experience and schooling are correlated with greater levels of productivity emanating from the unobservables, the remaining workers, particularly males, who are less educated and younger will receive lower returns from these worker characteristics. Since the model is overidentified, tests for the validity of identifying restrictions in the wage equation were performed. These tests were done by adding more regressors into the wage estimates and then, if found to be significantly non-zero, by determining whether the additional regressors are independent of the error term. In particular, parents’ education Žwhich may capture unobserved human capital. and the cohort dummies Žwhich can reflect schooling quality differences across the years, if any. were incorporated separately into the alternative specifications of wage estimates. These tests revealed that only the cohort variables were significant Žsee Appendix A.. Moreover, using the Lagrange Multiplier ŽLM. test to determine whether these variables were dependent of the error terms, different degress of validity are shown for each gender. Given an LM statistic of 4.06 for the females, these restrictions in the estimates are not significant at 10% level, with 2 degress of freedom. This suggests that the cohort variables are correlated with the error term, and hence not valid to use as independent variables. 7 For the males, however, the LM statistic of 8.47 is significant at a 5% level. In any case, the results did not alter substantially. While the estimated returns to education are lower, after accounting for the plausibly inferior quality of schooling in the previous years, the wage estimates still feature negative selectivity as well as substantially higher returns to education than those found in the OLS equations. Finally, the results show significant differences in the wages of men and women in both the OLS and the selection-corrected estimates. To determine if the coefficients for the male and female wage estimates are significantly different, Wald tests are performed ŽGreene, 1993.. 8 The tests obtained values of 39.87 for the OLS estimates and 123.68 for the selection-corrected estimates. The results are significant at a 0.05% level for both the OLS and the selectivity-corrected models with 5 and 6 degrees of freedom, respectively. To determine how much of this observed difference can be explained by the selectivity variable, the Oaxaca Ž1974. decomposition of wage differentials can be used. Table 6 reports the results of the decomposition of differences in wages
7 The exact nature of the dependence of these cohort variables with the error terms cannot however be known. For one, such variables may be poor indicators of schooling quality, thereby making the estimates subject to measurement errors. 8 A standard test of structural break is the F-test. However, the test assumes that the disturbance variances for both regressions are equal. This assumption is improbable because the variances for different samples may also be different ŽDavidson and MacKinnon, 1993.. If the model is heteroskedastic, results for classical regressions will no longer be valid.
46
Worker characteristics and selectivity control
Experience Years of schooling Selectivity control Log wages Gross differential Žin log wages.
Mean value
OLS estimates
Selection-corrected estimates
Male
Female
Effect of differences in variables
Share in gross differential
Effect of differences in variables
Share in gross differential
11.267 7.636 1.452 4.285
10.863 9.020 1.744 4.112
0.0192 y0.1040
0.1106 y0.6003
0.0277 y0.1236 0.1364
0.1600 y0.7132 0.7871
0.1733
0.1733
The effect of differences in variables is computed by taking the estimated coefficient of the variable from the men wage regressions and multiplying this by the differences in the mean values of the same regressor for men and women.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
Table 6 Decomposition of male–female wage differential
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
47
received by men and women using the estimated OLS and selection-corrected wage functions. 9 The objective of the Oaxaca decomposition is to evaluate how much of the male–female wage differential can be eliminated if women received the same human capital characteristics as men, assuming that they are rewarded on the basis of their estimated wage function. The selectvity-control variable is then presumed to capture the effects of identifying factors that influence labor market participation and locational choice. This exercise leads to the following conclusions. First, the OLS estimates appear to have limited capability in explaining the gender differences in wages, since controlling for education and experience does not narrow the 17% gap. This residual may include the effects of various unobserved variables, e.g., discrimination. Including education, experience and the selectivity term, the selection-corrected estimates are able to account for more than two-thirds of the positive male–female wage differential. Second, the means of the schooling variable indicate that women have more schooling than men, thereby substantially increasing the controlled wage differential. The results suggest that if women were brought down to the schooling levels of men, they will be even less well-off. Third and most important, the selectivity control variable singly accounts for 79 percentage points of the wage differential. The men’s greater likelihood of engaging in the wage labor market incorporates within it unobserved factors that result in greater wage differences between the sexes. Considering that the females led the men in outmigration, the outmigration of the more productive women from the rural areas seems to substantially contribute to the observed wage gap in these areas.
6. Conclusion The purpose of this paper was to determine if the selectivity of migration that leads to sample attrition affects the market wage structure estimated for those who stay in rural low-income economies. A model of wage determination was devel9
Assuming error terms are averaged out, the estimated wage equations can be used to account for differences in the logarithm of wages for men and women, as shown below ŽOaxaca, 1974.: Wm yWf s w Ž a m y af . qÝ Xf Ž bm y bf . x q w Ýbm Ž X m y X f . x where a’s are estimated intercepts, b’s the coefficients on the workers’ characteristics as well as other unobserved endowments as men, and the summation denotes the total effect of these factors. The last terms indicate the differences in wages that are explained by the differences in workers’ characteristics, weighted by the regression coefficients of the male sample. These then account for the effects solely of the differences in worker characteristics, while the first two terms may be simply viewed as unexplained differences. The effects of the differences in characteristics on the wage differential can also be weighted by female regression coefficients. However, since males dominate the wage labor market, their characteristic weights are used in Table 6.
48
Constant Experience Experience squaredr100 Years of schooling Lambda Additonal Variables Parents’ Education Father Mother Year data collected 1978 1983 Adjusted R 2 F-test N LM Statistic
Males 3.6541)) Ž22.05. 0.0691)) Ž5.89. y0.0011)) Ž7.15. 0.0881)) Ž8.92. y0.4809)) Ž12.04.
3.4378)) Ž17.92. 0.0795)) Ž5.58. y0.0013)) Ž7.23. 0.0867)) Ž10.99. y0.2501)) Ž5.31.
Females 4.1102)) Ž18.75. 0.0943)) Ž9.98. y0.0010)) Ž8.36. 0.1074)) Ž9.75. y0.9172)) Ž9.56.
4.2187)) Ž22.65. 0.0884)) Ž6.03. y0.0012)) Ž6.28. 0.0894))Ž9.54. y0.9736)) Ž19.41.
y0.0063 Ž0.50. 0.0214 Ž1.61.
y y
0.0198 Ž1.51. y0.0104 Ž0.76.
y y
y y 0.17 7.27 187 1.80
y0.2883)) Ž2.26. y0.3826)) Ž2.96. 0.19 8.31 187 8.47
y y 0.32 8.88 102 5.40
0.4102)) Ž3.51. y0.1576)) Ž1.91. 0.33 9.30 102 4.06
See notes in Table 5. The Lagrange Multiplier ŽLM. Statistic determines the independence of the addtional variables to the error term.
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
Appendix A. Alternative specifications for wage estimates
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
49
oped to account for self-selection due to migration. The main concern is whether or not the individuals with more ability and other unmeasured endowments tend to migrate out of the villages, leaving the less able in the place of origin. It is hypothesized that migration can result from a cost-benefit differential from locating in a particular region. The probability of observing a wage is conditional not only on the choice to participate in a wage market but also on the choice not to migrate to areas where amenities, such as educational services, are available at a lower cost. An analysis of Philippine data gathered indicates that more educated and experienced individuals are more likely to migrate. Moreover, there exists a negative selection bias in the wage estimates, resulting in lower wages for workers who remain in their parental home compared to the population mean conditional on observables. Because of this, wage returns to education and experience differ between the simple OLS estimates and the sample selection-corrected estimates. Differences in wage rates between men and women are also explained more satisfactorily in the estimates purged of selectivity. Since self-selection is shown to explain the observed wage differences between men and women, omitting this selection control variable is also expected to result in biased evaluations of human capital investments, specifically those directed to women. Persons remaining in their respective villages and parental homes tend to be less productive than those who have migrated. This finding is relevant to the issues of income distribution and public finance. A larger share of the costs of rural schooling should be subsidized by the national government which captures the gains nationally of a more productive work force.
Acknowledgements The author wishes to thank T. Paul Schultz, Robert Evenson, T.N. Srinivasan, and Michael Alba for their advice and comments at various stages of the study. The referees and editorial board of this journal contributed greatly to this present version. Most of the research was conducted at the Yale University Economic Growth Center, with the support of the Rockefeller Foundation. Remaining errors are the author’s sole responsibility.
References Anderson, K., 1982. The sensitivity of wage elasticities to selection bias and the assumption of normality. J. Hum. Resour. 17, 594–605. Becker, G., 1981. A Treatise on the Family. Harvard Univ. Press, Cambridge, MA. Becker, G., 1993. Human Capital, 3rd edn. University of Chicago Press, Chicago. Behrman, J., Birdsall, N., 1988. The equity-productivity trade-off: public school resources in Brazil. Eur. Econ. Rev. 32, 1585–1602.
50
L.A. Lanzonar Journal of DeÕelopment Economics 56 (1998) 27–50
Chiswick, C., 1976. On estimating earnings functions for LDCs. J. Dev. Econ. 4, 67–78. Davidson, R., MacKinnon, J., 1993. Estimation and inferences in Econometrics. Oxford Univ. Press, Oxford. Evenson, R., Binswanger, H., 1984. Estimating labor demand from Indian Agriculture. In: Binswanger, H., Rosenzweig, M. ŽEds.., Contractual Arrangements, Employment and Wages in Rural Labor Markets in Asia. Yale Univ. Press, New Haven, CT. Falaris, E., 1987. A nested logit migration model with selectivity. Int. Econ. Rev. 28, 429–443. Fields, G., 1975. Rural–urban migration, urban unemployment and underemployment, and job search in LDCs. J. Dev. Econ. 2, 165–187. Greene, W., 1993. Econometric analysis. Prentice-Hall, New York. Greenwood, M., 1971. An analysis of the determinants of internal labor mobility. Ann. Regional Sci. 5, 137–151. Harris, J., Sabot, R., 1982. Urban unemployment in LDCs: towards a more general search approach. In: Sabot, R. ŽEd.., Migration and the Labor Force. Westview Press, Boulder, CO. Harris, J., Todaro, M., 1970. Migration, unemployment and development: a two-sector approach. Am. Econ. Rev. 60, 126–142. Hausman, J., McFadden, D., 1984. Specification tests for the multinomial Logit model. Econometrica 52, 1219–1240. Heckman, J., 1979. Sample bias as a specification error. Econometrica 47, 153–161. Lee, L.F., 1983. Generalized econometric models with selectivity. Econometrica 51, 507–512. Levy, M., Wadycki, W., 1974. Education and the decision to migrate: an econometric analysis of migration. Econometrica 42, 377–388. Maddala, G., 1983. Limited Dependent and Qualitative Variables in Econometrics. Cambridge Univ. Press, Cambridge. McFadden, D., 1974. The measurement of urban travel demand. J. Public Econ. 3, 303–328. Mincer, J., 1978. Family migrations. J. Pol. Econ. 86, 749–773. Nakosteen, R., Zimmer, M., 1980. Migration and income: the question of self-selection. Southern Econ. J. 46, 840–851. Oaxaca, R., 1974. Male–female wage differentials in urban labor markets. Int. Econ. Rev. 14, 693–709. Rosenzweig, M., Wolpin, K., 1986. Evaluating the effects of optimally distributed public programs: child health and family planning interventions. Am. Econ. Rev. 76, 470–482. Rosenzweig, M., Wolpin, K., 1988. Migration selectivity and the effects of public programs. J. Public Econ. 37, 265–289. Schultz, T.P., 1988. Heterogeneous preferences and migration: self-selection, regional prices and programs, and the behavior of migrants in Colombia. In: Res. Pop. Econ., Vol. 6, JAI Press, CT, pp. 163–181. Schultz, T.P., 1982. Notes on the estimation of migration functions. In: Sabot, R. ŽEd.., Migration and the Labor Force. Westview Press Boulder, CO. Schwartz, A., 1973. Interpreting the effect of distance on migration. J. Pol. Econ. 81, 1153–1169. Singh, I., Squire, L., Strauss, J., 1986. Agricultural Household Models: Extensions, Applications and Policy. John Hopkins Univ. Press, Baltimore. Sjaastad, L., 1962. The costs and returns of human migration. J. Pol. Econ. 70, 80–93. Train, K., 1993. Qualitative Choice Analysis: Theory, Econometrics and an Application to Automotive Demand. MIT Press, MA. Willis, R., 1986. Wage determinants: a survey and reinterpretation of human capital earnings. In: Ashenfelter, O., Layard, R. ŽEds.., Handbook of Labor Economics, Vol. I. North-Holland, Amsterdam.