Discussion Papers In Economics And Business

Transcription

1 Dicuion Paper In Economic And Buine Self-election and Earning of Migrant: Evidence from Rural China Zheren WU Dicuion Paper Graduate School of Economic and Oaka School of International Public Policy (OSIPP) Oaka Univerity, Toyonaka, Oaka , JAPAN

2 Self-election and Earning of Migrant: Evidence from Rural China Zheren WU Dicuion Paper June 2008 この研究は 大学院経済学研究科 経済学部記念事業 基金より援助を受けた 記して感謝する Graduate School of Economic and Oaka School of International Public Policy (OSIPP) Oaka Univerity, Toyonaka, Oaka , JAPAN

3 Self-election and Earning of Migrant: Evidence from Rural China * Zheren Wu Abtract Uing data from a rural houehold urvey in China, thi paper explore the link between employment choice (nonworking, local farm work, local nonfarm work and migratory work) and migrant earning. We find ignificant elf-election in migration. Youth, male, better-educated individual and thoe in good health are more likely to migrate. In term of unoberved characteritic, we find poitive election in migration a related to the alternative of not working and local farm work, and inignificant elf-election a related to local nonfarm work. Controlling for elf-election, the wage return to gender (male), education and health are lower than thoe obtained from imple ordinary leat quare (OLS), and the return to experience are higher. More importantly, we find different elf-election between individual who have moved a pioneer and migrant from houehold in which other member have already migrated. JEL claification: J24; J31; O15 Keyword: Migration, Self-election, Pioneer migrant * Thi paper ha benefited from helpful comment provided by Tunehiro Otuki, Akira Kouaka, Hiakazu Matuhige, Miki Kohara, Shinya Kaitani and Junyi Shen. I am grateful for financial upport from the Matuhita International Foundation and the Setutaro Kobayahi Memorial Fund. Correpondence: Oaka School of International Public Policy, Oaka Univerity; 1-31 Machikaneyama, Toyonaka, Oaka, , Japan wzheren@oipp.oaka-u.ac.p

4 1. Introduction Migration ha long been recognized a a elf-election proce. Baed on the inight of Roy (1951), Bora (1987, 1991) invetigated how the endogeneity of the migration deciion affect labor market performance, in term of both unobervable characteritic (uch a ability, ambition, phyical trength and kill) and obervable characteritic (uch a education, experience, and training). It i expected that individual compare their potential wage (baed on their obervable and unobervable human capital) in their home community with thoe in other region. Thoe who migrate have higher expected earning in other region than at home and vice vera for thoe who tay. When migration i an endogenou outcome of optimizing deciion, wage etimate are inconitent unle corrective meaure are taken. Thi, in turn, will lead to mileading evaluation of policy program uch a public invetment in education and ubidie on training (Roenzweig and Wolpin, 1986; Lanzona, 1998). The elf-election in migration alo generate prediction about how migrant compare with the population of the origin. The prediction of migrant electivity are important for policy making. For example, there i a riing concern that large-cale rural-to-urban migration might harm the rural development in China. If able worker move out of the village, the lack of competent worker will dampen the development of rural economie, cauing rural area to fall further behind urban area. In thi cae, local government may have an incentive to carry out policie to retrict migration. 3 Becaue migrant attribute uch a ability and kill are unobervable, policy maker are alway limited in obtaining predication for elf-election by uing obervable characteritic uch a education. However, there i no neceary correlation between the elf-election in term of obervable characteritic and in term of unobervable characteritic. For intance, it i perfectly poible for migrant to be among the mot educated in their home communitie but perform poorly in the detination (Ghatak et.al., 1996). Hence, invetigating elf-election baed on unoberved characteritic will yield important implication. Heckman (1976, 1979) developed a two-tep procedure to correct elf-election bia and infer the unoberved characteritic by checking whether individual unoberved characteritic from their choice of whether or not to participate in the activity of interet (e.g., migration) are correlated with the outcome of interet (e.g., migrant earning). Lee (1983) and Dubin and McFadden (1984) developed the approach to apply to multiple choice. In general, when the unobervable characteritic from the migration deciion are conitently aociated with a higher (lower) wage rate, we conclude migrant are more (le) productive and poitively (negatively) elected. Although the exiting literature ha provided deep theoretical inight into, and variou econometric methodologie toward undertanding, the elf-election in migration, empirical tudie to acertain the impact of unobervable characteritic on the migration deciion and earning etimation are till few in number. It might be becaue information (epecially regarding wage) of migrant i alway miing from urvey conducted in their origin. A an alternative, Lanzona (1998) analyzed whether the electivity of 3 The effect of migration on the economic of ource community alo depend on remittance and other factor. See, for example, Rozelle, et.al. (1999), de Brauw and Rozelle (forthcoming), Ghatak et.al. (1996)

5 migration affect the wage tructure etimated for thoe who tay in rural communitie. Uing Philippine data and Lee ample election model, Lanzona found that there i a negative election bia in nonmigrant wage etimate, and that a a reult migrant were poitively elected. Axelon and Weterlund (1998) examined the potential influence of migration on total houehold real income in Sweden. Employing Heckman election bia correction approach, they accounted for the potential correlation between unoberved houehold characteritic exerting influence on both the deciion to migrate and on houehold income but did not find ignificant elf-election in migration. In thi paper, we ue data from a rural houehold urvey conducted by the Minitry of Agriculture of China in the Sichuan and Anhui province to explore the link between employment choice (nonworking, local farm work, local nonfarm work and migratory work) and migrant earning. It i worth pointing out that, under the unique houehold regitration ytem in China (ee Liu, 2005; Au and Henderon, 2006; etc.), entire rural houehold rarely relocate, and migrant uually maintain their permanent regitration in the home and circulate between their place of origin and detination. Thee feature enable the rural houehold urvey to collect data uccefully on every houehold member, including migrant. Therefore, miing obervation for migrant in the ample, which i a typical problem in other countrie, doe not arie. In contrat to the traditional approach in exiting tudie (typically a verion of Heckman, 1976, 1979; or Lee, 1983), we employ the electivity correction methodology developed by Dubin and McFadden (1984) and modified by Bourguignon et al. (2007). Thi ophiticated method, baed on the multinomial logit model including multiple correction term, allow u not only to attribute a election bia in the etimation of earning to the allocation of individual with better or wore unoberved characteritic in migration, but alo to link the election bia to the allocation of individual to each other alternative. That i, we are allowed not only to identify what the election bia i, but alo from where the bia tem. The empirical etimate document ignificant elf-election in migration. In general, youth, male, better-educated individual and thoe in good health are more likely to migrate. In term of unoberved characteritic, we find poitive election in migration a related to the alternative of not working and local farm work, and inignificant elf-election a related to local nonfarm work. Controlling for elf-election, the wage return to gender (male), education and health are lower than that obtained from OLS, and the return to experience are higher. More interetingly, when we conider the poibility that pioneer migrant may confer a poitive externality on (potential) future migrant, the empirical reult baed on two ubgroup, which are divided by whether or not other houehold member had moved in the pat three year, exhibit triking difference. Thi ugget that the elf-election in migration i tied ignificantly to other houehold member migration experience. Thi paper i tructured a follow. Section 2 preent the econometric framework. Section 3 decribe the data and the empirical pecification. Section 4 report the empirical reult and Section 5 conclude

6 2. Econometric Model: Selection Bia Correction Baed on Multiple Choice Conider a ituation in which an individual make an employment choice, each individual may elect among J mutually excluive alternative. Thee alternative could for example be: (a) not to work; (b) to be a local farm worker; (c) to be a local nonfarm worker; and (d) to migrate from the home village and work in ome other location. Let Y * be the utility attainable for an individual if he/he chooe alternative. We can write the indirect utility function a: Y * = Z + η, γ = 1,..., J, (1) where the vector Z repreent the maximum et of explanatory variable for all employment alternative, and η i a diturbance term. A rational individual compare the utility attainable given each alternative and elect the alternative that yield the highet benefit to him/her, that i: Y * > max( Y * ), (1,..., J ). Aume the market wage in the th alternative i given by: ln w = X β + u, (2) where X i a vector of exogenou variable that determine logarithmic earning (ln w ), and the diturbance u atifie E( u X, Z) = 0 and V 2 ( u X, Z) σ =. If there are unoberved characteritic that affect both individual choice and their earning, then the diturbance η in (1) and diturbance u in (2) will be correlated uch that: E(ln w Y * > maxy * ) = X β + E( u = X β + E( u X β. Y * > maxy * max( Z γ + η Z γ ) η ) < 0) The term E( u max( Z + Z ) < 0) capture the unobervable characteritic affecting γ η γ η deciion making and are peritently correlated with the wage. For intance, able individual could chooe to migrate, and potentially reward their ability with a higher wage rate. Becaue we cannot oberve ability, omitting important variable from the OLS regreion produce biaed etimate of β. The potential inconitency require election bia correction method, following the eminal inight of Heckman (1976, 1979). When the choice are multiple, two typical approache baed on the multinomial logit pecification were developed by Lee (1983) and Dubin and McFadden (1984, hereafter DMF). Lee - 3 -

7 model i baed on the aumption that the unobervable determinant of the choice of alternative againt any other alternative correlate with the unobervable determinant of the outcome ( ln ) in the ame direction, and thu the model ue only one correction term to explain E( u max( Z γ + η Z γ η ) < 0). DMF make no aumption about thi. They ue multiple th correction term to control for elf-election in the alternative a related to each other alternative. The correlation between u and η η ) could be of different ign for different. A hown by ( Schmertmann (1994) and Bourguignon et al. (2007), Lee avoid the rik of multicolinearity preent in DMF but make a very trong aumption that doe not alway hold in empirical tudie. If Lee aumption did not hold, then the reult would be inconitent. Another virtue of DMF approach i that it identifie not only the direction of the election bia, but alo where the bia tem from, by linking the election bia to the allocation of individual to each other alternative. The formulation of the ample election model in thi paper follow DMF and here we briefly outline DMF method. A more detailed dicuion can be found in Dubin and McFadden (1984) and Bourguignon et al. (2007). Auming the diturbance η in equation (1) are independent and identically Gumbel ditributed (the o-called IIA hypothei), the pecification of (1) lead to the multinomial logit model with: w P k = exp( Zγ ) J = 1 k exp( Zγ ), = 1,...,J, k (1,..., J ), (3) where P k i the probability that any alternative k i preferred. Under DMF aumption that: E 6 ) = σ r ( η E( η )), with r = 0, 1,..., J, π ( u = conditional on the alternative being choen, etimating the following wage function that include the correction term yield conitent etimate of β : ln w 6 P ln( P ) = X β + σ r[ + ln(p )] π 1 P + e, (4) where r i the correlation coefficient between diturbance aymptotic mean zero. u and η, and e i a reidual with Bourguignon et al. (2007) howed that when we remove the retriction of r = 0, equation (4) can be written a: - 4 -

8 ln w 6 = X β + σ [ r π P ln( P ) r ln(p )] 1 P + e. (5) Their Monte Carlo experiment howed that the equation (5) variant outperform the traditional DMF correction method when the retriction i violated and provide very imilar reult when it hold. Equation (4) and (5) can be etimated in two tep. In the firt tep, the polychotomou choice model i etimated by the logit maximum likelihood method (equation (3)). Let Pˆ be the predicted probabilitie for P, = 1,..., J. In the econd tep, we ubtitute Pˆ, = 1,..., J into the electivity correction term in equation (4) or (5) and we then etimate the function by OLS. Note that while the econd tep etimate from DMF are conitent, they have inefficient tandard error becaue of the two-tep nature of the procedure. There can be ome efficiency gain from uing weighted etimation. The form of heterokedaticity i detailed in Bourguignon et al. (2007) and we ue the boottrap method. Becaue σ 0, we conduct a t -tet of whether 6 σ r π = 0 can be ued to tet r = 0. The imple OLS covariance matrix i valid only when all the elf-election correction coefficient equal zero. Furthermore, we notice that the value of P ln( P ) increae with the decreae in 1 P P. Then a poitive (negative) coefficient of 6 σ π indicate a poitive (negative) election becaue the individual with r better (wore) unoberved endowment chooe the given alternative rather than the repective alternative. For example, a poitive elf-election correction coefficient related to local farm work in the migrant wage equation highlight the higher wage of individual who work outide the home village compared with random individual, becaue people who are more productive (in term of unoberved characteritic) prefer migration to local farm work. 3. Data and Model Specification 3.1 The data The empirical analyi i baed on the data from the rural houehold urvey conducted by the Minitry of Agriculture of China (MAC) at the end of December each year. We ue data on the Sichuan and Anhui province that cover the period from 2003 to The Sichuan and Anhui province are in wetern and central China, repectively. Both are predominantly rural and are the two larget exporter of rural labor. Thu, the experience of rural houehold in thee two province may hed light on thoe of rural houehold elewhere in le developed region in China. The urvey include approximately 2000 houehold from 33 village. The original election of village and houehold ued tratified random ampling method. From 2003, the MAC collected data not only at the houehold level but alo on each houehold member. Houehold-level data include information on - 5 -

9 income, landholding and other apect of houehold economic activity. Data on individual include information about employment, education and demographic characteritic. An outtanding feature of our data i that, under the unique houehold regitration ytem, entire rural houehold rarely relocate; migrant uually maintain their permanent regitration in the home village and circulate between their place of origin and detination. Therefore, another type of ample election, i.e., the omiion from the urvey of migrant, which i a problem that typically arie in urvey in mot other countrie, doe not arie in China. One of the hortcoming of the data i that peronal earning are only available in the migration ection, thu we can focu on the migrant wage equation only. Becaue mot migrant moved repeatedly, our analyi mainly focue on 2006 data, however utilize information uch a migration experience in the pat three year ( ). We include only laborer in the ample. Following the definition by the National Statitical Bureau of China, a laborer i defined a a female aged between 16 and 55 or a male aged between 16 and 60. Anyone who work 90 day or more during the year but fail to meet the age contraint alo belong to the labor force. 4 Full-time tudent and individual who had lot the ability to work are excluded. After accounting for miing obervation, the valid ample conit of 4,820 individual from 1,797 houehold. We claify thee rural laborer into four categorie: (1) nonworking individual; (2) local farm worker; (3) local nonfarm worker; and (4) out-of-village migrant. Anyone who worked le than 90 day in the labor market in a year i defined a a nonworking individual. 5 A migrant i defined a a peron who worked outide hi or her home village on at leat 30 day during the urvey year. 6 Under thi definition, 2,121 individual in our ample migrated and 94.15% of them worked outide their village on at leat 90 day; the mean migration period i 254 day. The ret of the population i divided into local farm worker and local nonfarm worker by elf-reported main ob. Table 1 preent the allocation of the labor force over the four election categorie and alo by gender, age and chooling cohort. A mentioned above, the Sichuan and Anhui province are agricultural province and large migrant-exporter; 39.63% of the ample population mainly participated in local farm work, and 44% of the population engaged in migratory work. Male appear to lead female in moving out the village (20.84% higher), however had a ignificantly lower probability of being a local farm worker (23.37% lower). We oberve a perceptible effect of age on the employment choice. Younger peron have a higher rate of moving out. Around 70% of thoe under 31 year old participated in out-of-village work. Older peron are more likely to be involved in local work. In the five age categorie, the highet participation rate in local farm and local nonfarm work both appeared in the cohort of over 35 year old, with rate of 54.71% and 10.83%, repectively. It may be becaue local work can eaily be combined with living in the village and old peron are likely to have more experience and more contact that are relevant for finding local nonfarm ob. Another notable point i that age eem negatively related to the proportion of unemployment, with 15.55% in the cohort of year old and 5.81% in thoe more than 35 year old. 4 When we et the minimum number of day a 1 or 30, our reult were imilar. 5 When we et the minimum number of day a 1 or 30, our reult were imilar. 6 We alo et the minimum number of day of migratory work a 1 or 90. The reult were imilar

10 Table 1 Employment election: by gender, age and education cohort (2006 MAC urvey) Employment categorie Total (people) (1)Nonworking (2)Local farm (3)Local nonfarm (4)Migration Total 4, % 39.63% 8.76% 44% Gender Male 2, % 29.03% 11.12% 53.74% Female 2, % 52.40% 5.90% 32.90% Age categorie % 8.54% 3.05% 72.87% % 11.21% 6.62% 72.24% % 15.91% 6.06% 68.75% % 28.52% 5.56% 57.23% 36-2, % 54.71% 10.83% 28.65% Education categorie 0-6 2, % 62% 7.50% 24.45% 7-9 2, % 22.83% 8.76% 60.75% % 9.39% 15.26% 59.86% The table alo how ignificant difference in the employment choice among people with different education endowment. Three main point can be noted. Firt, migration mainly occurred among better-educated people. The proportion i around 60% for thoe who had at leat even year of chooling. Second, le-educated people are more likely to engage in local farm work. Of thoe belonging to the 0 6 year of chooling cohort, 62% engaged in local farm work, far exceeding the participation rate of 22.83% and 9.39% for thoe belonging to the 7 9 year of chooling cohort and more than 9 year of chooling cohort, repectively. Third, compared with le-educated individual, better-educated people are more likely to be unemployed. We find that 15.49% of individual who received at leat 10 year of chooling did not participate in the labor market. Educated unemployment ha been a noteworthy phenomenon in rural China. 3.2 Empirical model and variable The two-tage approach require that there mut be different et of independent variable in the two function. Following the preceding tudie (ee, for example, the urvey of Greenwood, 1975, 1997; Yap, 1977; Luca, 1997), the empirical pecification i defined a: P = f ( Network, Migration experience, Age, Age quared, Male, Education, Health, Number of young dependent, Number of other migrant, Houehold ize, Landholding, Lagged houe value, Village dummie), = 1,2,3,4, (6) - 7 -

11 ln(daily earning) = g( Netwrok, Migration experience, Age, Age quared, Male, Education, Health, Number of young dependent, correction term), iff = 4, (7) where (6) i the employment election function and (7) i the migrant earning function. Table 2 preent definition of main variable and report their mean and tandard deviation for the whole ample and the four categorie. We ue logarithmic average daily migrant earning a a dependent variable in the earning function. The mean value of migrant earning per day i 28.8 yuan. From the dicuion above and the migration literature, we expect that thoe who have migration experience or migration network, male, better-educated individual and healthy people are more likely to move and meet with ucce at their detination. Note that the network dummy take the value 1 if any other houehold member had been a migrant, while the individual had not moved in the pat three year, and take the value 0 otherwie. Thi dummy indicate the preence of an effective migration network that provide information to potential migrant without previou migration experience. Older people are le likely to migrate, becaue they have le time to pay back invetment and they often have their family to take care of. However, we predict a poitive coefficient on age in the earning function, a age i an intrument of experience. The quared term i added to account for nonlinearitie in the impact of age. For the number of young dependent, individual with children are generally le eager to move, however the need for a large houe or the (propective) chooling of children might prompt a move (Sandell, 1977; Mincer, 1978; Nivalainen, 2004; etc.). In addition, people with dependent children are uually viewed a being more reponible. Therefore, the effect of the number of young dependent on the migration deciion might be ambiguou, while the effect on earning might be poitive. Variable of the number of other migrant, houehold ize, landholding, lagged houe value and village factor are hypotheized to influence the occupation deciion, however have no effect on the migrant wage in their detination. The effect of the number of other migrant on an individual migration deciion i unpredictable: when migration occur by one or more houehold member, other houehold member may chooe to move with them or to remain in the village to look after the home. Worker belonging to large houehold are expected to be le mobile becaue of their greater attachment to the family. Large area of irrigated land reduce the need for working off-farm and the poibility of unemployment. The variable of lagged houe value meaure wealth. People from poor houehold are expected to have an incentive to migrate becaue they are more motivated than the rich by potential income gain from migration. The highet average lagged houe value i found in thoe who worked in local nonfarm ector. A et of village dummie i included in (6) to control village characteritic uch a village infratructure, geographic location, economic hock becaue of natural diater, migration network with neighbor, and local wage level. Conitent with the tatitic in Table 1, Table 2 how that thoe who did not participate in the labor market are relatively young. Note alo that the average chooling year of thoe who are unemployed i ignificantly higher than that of local farm worker, and even lightly higher than that of local nonfarm - 8 -

12 Table 2 Definition and decriptive tatitic (mean) of elected variable Total (1) (2) (3) (4) Definition Nonworking Local farm Local nonfarm Migration Daily earning average daily earning, (yuan) (16.706) #Network dummy,=1 if other member had moved in and the individual had not move in (0.471) (0.501) (0.496) (0.462) (0.300) #Migration experience dummy,=1 if the individual had moved in (0.500) (0.444) (0.342) (0.438) (0.331) Age year old (13.898) (13.516) (12.044) (13.638) (11.694) #Male dummy,=1 if male (0.498) (0.497) (0.490) (0.461) (0.471) Education year of chooling (2.990) (3.446) (2.806) (3.117) (2.340) Health =5, very good; =4 good; =3 normal; =2, bad; =1, very bad. (0.783) (0.808) (0.843) (0.880) (0.643) Number of young dependent Number of prechool children and tudent (people) (0.807) (0.747) (0.859) (0.862) (0.756) Number of other migrant (people) (0.988) (0.989) (0.997) (0.878) (0.977) Houehold ize (people) (1.530) (1.537) (1.660) (1.679) (1.359) Landholding Per capita landholding (mu) (4.601) (4.580) (5.109) (4.175) (4.124) Lagged houe value value of houe in 2005 (1,000 yuan) (34.575) (70.181) (26.089) (37.568) (30.648) N 4, , ,121 Note: Standard error in bracket; # indicate dummy variable

13 worker. Why did young educated people not participate in the labor market? A poible explanation for thi phenomenon i that thee individual reervation wage are relatively high, and they did not want to work a a farmer. If the opportunity of local nonfarm employment i limited or the wage are low, they cannot or will not enter the local nonfarm ector. However, why did they not work outide the village? A potential reaon might be that they fear an uncertain life in other location becaue of a lack of information. From our data, we find that for individual in the houehold in which other member had not moved in the pat three year, 33.85% of year-old young laborer and 20.98% of thoe who have received at leat 10 year of chooling were unemployed. The proportion are ignificantly lower for individual from houehold in which other member had migration experience, with 11.03% and 12.72%, repectively. According to Carrington et al. (1996) and McKenzie and Rapoport (2007), the preence of pioneer migrant may lower moving cot and enhance migration ob propect of other houehold member (potential future migrant). They increae the information available to other member and reduce the rik of moving by ending information about how to migrate, where to look for work, what wage to expect, and information on migration cot and rik and how to reduce them. Furthermore, if future migrant move to the ame region, pioneer migrant can upport thee ubequent migrant at their detination by, for example, providing them with ob earch aitance, helping them to find houing, and even by extending credit, providing lodging and in other way lowering the pychological cot of leaving home. Hence, individual from houehold in which other member have had a migration experience are more likely to migrate and le likely to be a nonworker. They are even le likely to be a local wage earner, if migration return are high. To addre thi iue, we divide the full ample into two ubgroup according to whether or not other member had moved in the pat three year, and ee whether choice making and the poible elf-election of migrant differ between thoe who moved a pioneer and thoe having acce to a network of previou migrant in the houehold Empirical Reult 4.1. Multinomial logit reult on employment election In the firt tep of our DMF analyi, we etimate the employment choice model a given in equation (6). Thi tep not only provide inight into the determinant of the choice of an individual, but alo generate bia correction term for the econd tep. Table 3 report the average marginal effect from the multinomial logit model for the full ample and the two ubgroup. In each cae, the tet baed on a chi-quare tatitic propoed by Hauman and McFadden (1984) doe not reect the null hypothei of IIA at the 99% ignificance level, indicating that ditinguihing between nonworking, local farm work, local nonfarm work and migration atifie the baic aumption in DMF. 7 Migrant from houehold where other member had not moved in the pat three year can be viewed a pioneer migrant. However, becaue of data limitation we cannot rule out the poibility that the migrant who come from houehold in which other member have migration experience were in fact the firt to move and therefore are pioneer. Neverthele, we invetigate the effect of prior migrant

14 Table 3 Marginal effect from multinomial logit for employment deciion Full ample Subgroup A: Other had NOT moved in Subgroup B: Other had moved in (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Nonworking Local farm Local nonfarm Migration Nonworking Local farm Local nonfarm Migration Nonworking Local farm Local nonfarm Migration #Network * ** ** (0.010) (0.015) (0.008) (0.016) #Migration experience ** ** ** ** ** ** ** ** ** ** ** ** (0.008) (0.019) (0.008) (0.026) (0.009) (0.026) (0.015) (0.033) (0.007) (0.014) (0.007) (0.017) Age ** ** ** ** ** ** ** ** ** ** (0.002) (0.002) (0.002) (0.002) (0.003) (0.004) (0.004) (0.004) (0.002) (0.003) (0.002) (0.003) Age quared/ ** ** ** ** ** ** ** ** (0.002) (0.003) (0.002) (0.003) (0.003) (0.005) (0.005) (0.004) (0.003) (0.003) (0.002) (0.004) #Male ** ** ** ** ** ** ** ** ** 0.03 ** ** (0.007) (0.011) (0.010) (0.011) (0.012) (0.022) (0.022) (0.020) (0.009) (0.013) (0.010) (0.013) Education ** ** * ** ** ** ** ** (0.001) (0.002) (0.001) (0.002) (0.002) (0.004) (0.003) (0.003) (0.002) (0.002) (0.002) (0.003) Health ** ** ** ** (0.005) (0.007) (0.006) (0.008) (0.010) (0.016) (0.014) (0.014) (0.007) (0.008) (0.006) (0.009) Number of young ** ** ** * ** 0.02 ** (0.006) (0.008) (0.006) (0.008) (0.009) (0.014) (0.012) (0.012) (0.008) (0.010) (0.007) (0.011) Number of other migrant ** ** ** ** ** ** 0.03 ** ** (0.005) (0.007) (0.006) (0.007) (0.019) (0.028) (0.026) (0.019) (0.005) (0.008) (0.005) (0.008) Houehold ize ** ** ** ** ** ** ** ** ** * (0.004) (0.006) (0.004) (0.005) (0.007) (0.010) (0.009) (0.009) (0.005) (0.007) (0.005) (0.007) Landholding ** ** ** ** ** * (0.001) (0.002) (0.002) (0.002) (0.002) (0.003) (0.004) (0.003) (0.002) (0.002) (0.002) (0.003) Lagged houe value ** * ** ** * ** ** (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 (0.000 Log likelihood LR tet [degree of ** [132] ** [129] ** [129] McFadden' Ad R N 4,820 1,399 3,421 Note: Standard error in bracket; ***, **, and * indicate ignificance at the 1%, 5%, and 10% level, repectively; # indicate dummy variable. Village dummie are all controlled

15 In the reult for the full ample in the firt four column, we find that both the individual and other houehold member migration experience (network) increae the probability of current migration. At the ame time, thee migration experience influence other alternative negatively, although the negative effect of network on the alternative of being a nonworking peron doe not have a high ignificance level. A ha been inferred from the decriptive tatitic, age ha a negative effect on the probability of leaving the home village. At the ame time, younger people how a greater propenity for unemployment. A for gender and education, male and better-educated individual are more likely to participate in local nonfarm work and migration. Health ha a poitive effect on the likelihood of migration and a negative effect on nonworking. The average marginal effect of other control variable are a expected, and are thu conitent with the theory and with exiting empirical finding. Column 5 to 12 in Table 3 preent the reult on the two ubgroup, with column 5 to 8 indicating individual from houehold in which other member had not moved in the pat three year (ubgroup A), and column 9 to 12 indicating thoe belonging to houehold in which one or more other member had moved (ubgroup B). The table reveal prominent difference between the two ubgroup. The mot noteworthy of the reult are a follow. Firt, for the choice of being a nonworker, a een in ubgroup A (column 5), without a network of prior migrant in the houehold, age, gender and education all ignificantly affect the likelihood of not working at leat at the 5% ignificance level. Generally, youth, female and better-educated individual are more likely to be unemployed. However, in ubgroup B, with at leat one other houehold member having migration experience (column 6), thee variable all failed to affect the likelihood of unemployment at the 10% tatitical ignificance level. There i no trong evidence to upport the hypothee that individual with different age, gender and education endowment have different probabilitie of unemployment. Second, a for the likelihood of being a local nonfarm worker, for individual belonging to houehold in which other member did not have migration experience (ubgroup A, column 8) there i an inverted U-haped effect of age; however for individual from houehold in which other member had migration experience (ubgroup B, column 11), the effect of age i horehoe haped. A poible explanation i that, becaue of the lack of network for individual in ubgroup A, local nonfarm work i a relatively favorable election, becaue older laborer have more experience and more contact that are relevant for finding local nonfarm ob, age increae the poibility of being a local nonfarm worker at firt. When people reach late middle-age, the effect of age become negative, indicating a link between nonfarm work and retirement. However becaue of the preence of migration network in ubgroup B, facing a lower moving cot and rik, youth, epecially thoe of marriageable age, are more eager to move out to make potential income gain from migration, and aving money for marriage or the arrival of children. When children go to chool, location tie are tronger and they may prefer to return to the village. In line with thi explanation, we oberved a poitive effect of age on the probability of local nonfarm work after individual reach their 30. Third, in ubgroup A, in the abence of a houehold member with migration experience, better-educated worker eem to chooe rural nonfarm work over migration. The average marginal effect of education on

16 the likelihood of chooing local nonfarm work i and ignificant at the 1% level, while it i on the likelihood of migration and inignificant at the 10% level. The null hypothei that education affect the likelihood of the two alternative equally i reected at the 10% ignificance level. In ubgroup B in which at leat one other houehold member had migrated in the pat three year, the average marginal effect of education on the two election are very imilar, reported by the ame value of 0.004, at the 5% ignificance level for local nonfarm work and at the 12% ignificance level for migration. The hypothei that education ha the ame effect on the two alternative i accepted at greater than the 95% level in the cae of ubgroup B. Fourth, with the abence of a pioneer migrant in the houehold, in ubgroup A, we fail to find ignificant effect attributable to health. Health i not an important determinant in individual employment deciion. On the other hand, in ubgroup B, for people belonging to houehold in which other member had experienced migration, the average marginal effect of health on the likelihood of migration i poitive (0.03), and the effect on the likelihood of nonworking i negative (-0.034); both are reported at the 1% ignificance level. People in good health are more likely to migrate and are le likely to be unemployed. Thee difference indicate that the preence of people who have experienced migration ignificantly change other member employment choice. Compared with thoe belonging to houehold where other member had experienced migration (ubgroup B), in houehold without a prior migrant (ubgroup A), health concern are not a main reaon of unemployment; youth and female are more likely to face unemployment; moreover, in the abence of uch a network, better-educated people are more likely to engage in local nonfarm work rather than migrate. If thee better-educated laborer fail to find nonfarm ob in the village, they prefer to wait and not work. From the etimation, the mean value of the predicted probability of unemployment in ubgroup A i 21.66% for individual between 16 to 20 year old, 10.15% for female, and 17.39% for thoe who received 10 or more year chooling; the correponding probability in ubgroup B i ignificantly lower 8 at 10.74%, 9.22% and 9.09%, repectively. Hence, we conclude that prior migrant play an active role in promoting migration and reducing youth unemployment and female unemployment in rural region; they alo enhance efficiency of the utilization of human reource by reducing educated unemployment Etimation of migrant earning In the econd tep, we etimate the migrant earning equation. To explore the importance of elf-election and highlight the difference between the OLS and the election-corrected wage etimate, we report both the OLS and DMF reult. Becaue the election bia correction baed on equation (5) are alway preferred to the approach baed on equation (4) (Bourguignon et al., 2007), and in our cae the reult of the two pecification were lightly different, we report the reult baed on equation (5) only. 9 8 Significant at the 1% 11% level in the one-ided t tet. 9 The Wu-Hauman tet cannot reect the null hypothei that there i no difference between the two pecification (without boottrap) at the 10% level. The reult baed on equation (4) are available upon requet

17 Table 4 preent the etimated reult for logarithmic average daily earning of migrant, where Table 4-2 include cro term of education and the dummy of migration experience into the pecification in Table 4-1. The etimator variance for DMF are all boottrapped with 100 replication to deal with the heterokedaticity. 10 In each DMF pecification, we find ignificant coefficient on the election bia correction term, which indicate that there i ignificant elf-election in migration and the imple OLS etimator are inconitent and DMF i preferred. 11 Uing the full ample, reult contained in column 4 how that the coefficient on the election bia correction term related to the nonworking ector and the local farm ector are both poitive and ignificant at the 5% level. In other word, migrant earning are upward biaed becaue the individual with better unoberved characteritic are more likely to chooe migration rather than to be unemployed or to work in the local farm ector. The coefficient on the correction term a related to local nonfarm work i negative but inignificant at the 10% level, indicating that there i no ignificant difference in productivity baed on unoberved characteritic between migrant and local nonfarm worker. Confirming the elf-election in the two ubgroup (column 5 and 6), in each cae we find a poitive coefficient on the correction term a related to local farm work (at leat at the 10% ignificance level). Similar to the reult for the full ample, there i an upward bia in migrant earning becaue of people who are well-endowed in term of unoberved characteritic prefer migration to local farm work. However, a oberved in the coefficient on the other two correction term, whether other houehold member do or do not have migration experience make a great difference. Firt, conider the correction term related to the nonworking ector. The coefficient i mall (0.09) and tatitically inignificant at the 10% level in ubgroup A (other member without migration experience), while the correponding coefficient i and ignificant at the 1% level in ubgroup B (other member had moved in the pat three year). Another difference i een in the coefficient on the correction term related to the local nonfarm ector. The coefficient i negative ( 0.377) and tatitically ignificant at the 5% level for migrant belonging to ubgroup A; 12 however, the coefficient i poitive and inignificant at the 10% level for thoe belonging to ubgroup B. In the abence of any other houehold member who migrated (ubgroup A), there i no ignificant elf-election in migration a related to unemployment, but negative election a related to local nonfarm work. In other word, individual with the bet unoberved characteritic prefer local nonfarm work; thoe with the wort endowment chooe to work in the local farm ector; the remaining worker who lie in the middle-level of unoberved endowment chooe migration or unemployment. However, in the preence of a member with migration experience (ubgroup B), migrant are poitively elected relative to unemployed individual; and tatitically inignificantly elected relative to local nonfarm worker. That i, the local nonfarm ector and migration ector attract worker with higher productivity, while le able 10 We alo repeated the boottrap algorithm 50, 200 and 1000 time. The reult were imilar. 11 The Wu-Hauman tet reect all the null hypothee of no difference between the etimate of OLS and DMF (without boottrap) at the 10% ignificance level. 12 Lee aumption do not hold in the cae of ubgroup A becaue the direction of elf-election in migration a related to local farm work and that a related to local nonfarm work are different

18 Table 4 Etimated earning equation reult OLS DMF OLS DMF (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Full Subgroup A Subgroup B Full Subgroup A Subgroup B Full Subgroup A Subgroup B Full Subgroup A Subgroup B #Network (0.079) (0.092) (0.079) (0.093) #Migration experience *** ** * (0.072) (0.078) (0.039) (0.110) (0.150) (0.075) (0.100) (0.280) (0.110) (0.190) (0.280) (0.150) Age *** *** *** *** *** *** *** *** *** *** *** *** (0.006) (0.013) (0.007) (0.006) (0.011) (0.008) (0.006) (0.013) (0.007) (0.007) (0.014) (0.007) Age quared/ *** *** *** *** *** *** *** *** *** *** *** *** (0.008) (0.016) (0.010) (0.007) (0.012) (0.010) (0.008) (0.016) (0.010) (0.008) (0.015) (0.009) #Male *** *** *** *** *** *** *** *** *** *** *** *** (0.025) (0.071) (0.027) (0.024) (0.059) (0.030) (0.025) (0.071) (0.027) (0.030) (0.061) (0.034) Education *** *** ** ** * (0.005) (0.012) (0.006) (0.006) (0.014) (0.006) Education*New (0.011) (0.033) (0.012) (0.012) (0.033) (0.013) Education*Repeat *** *** ** ** ** * (0.006) (0.012) (0.007) (0.007) (0.015) (0.007) Health *** *** *** *** *** *** *** ** *** *** ** *** (0.018) (0.040) (0.021) (0.021) (0.037) (0.022) (0.018) (0.040) (0.021) (0.019) (0.045) (0.021) Number of young dependent *** ** *** *** ** *** *** ** *** *** ** *** (0.016) (0.033) (0.018) (0.014) (0.037) (0.018) (0.016) (0.033) (0.018) (0.017) (0.033) (0.019) DMF1 (related to nonworking) *** *** *** *** (0.120) (0.180) (0.130) (0.097) (0.200) (0.110) DMF2 (related to local farm work) ** * * ** * * (0.130) (0.190) (0.140) (0.120) (0.210) (0.160) DMF3 (related to local nonfarm work) ** ** (0.120) (0.170) (0.130) (0.130) (0.220) (0.120) DMF4 (related to migration) (0.037) (0.066) (0.047) (0.048) (0.069) (0.053) Contant *** *** *** *** *** *** *** *** *** *** *** *** (0.160) (0.320) (0.160) (0.200) (0.350) (0.180) (0.160) (0.410) (0.180) (0.220) (0.430) (0.170) Aduted R-quared N 2, ,654 2, ,654 2, ,654 2, ,654 Note: Standard error in bracket; ***, **, and * indicate ignificance at the 1%, 5%, and 10% level, repectively; # indicate dummy variable. Subgroup A refer individual who come from houehold in which other member had not moved between 2003 and 2005; Subgroup B refer individual who come from houehold in which other member had moved between 2003 and 2005; DMF, =1,2,3,4 i the coefficient on correction term of P lnp/(1- P ), = 1,2,3and lnp, repectively; 100 boottrap replication are performed in DMF

19 worker either do not work or are located in the local farm ector. The different elf-election in the two ubgroup indicate that employment choice i ignificantly tied to other member migration experience. The preence of uch a migration network promote an efficient allocation of labor reource in that people with better unoberved characteritic flow out of the nonworking ector, or perhap even the local nonfarm ector, to the migration ector. The difference in the etimated coefficient between OLS and DMF are equally intereting. In each cae, OLS overetimate the gender difference and the return to education. Thi i the reult of the poitive election in migration in term of gender (male), education and unoberved characteritic. The difference on the return to experience (age) and health between OLS and DMF provide another et of intereting obervation. OLS underetimate the return to experience and overetimate the return to health for the full ample and ubgroup B. It i undertandable: migrant are negatively elected in term of experience (age) and poitively elected in health, becaue the elf-election baed on unoberved characteritic i poitive on average. Thee factor lead to a downward bia in the etimator of the return to experience (age) and an upward bia in the return to health when we fail to correct for elf-election bia. However, for ubgroup A, OLS lightly overetimate the return to experience and underetimate the return to health. Recall that both experience and health affect the likelihood of migration and local nonfarm work in the oppoite direction. The obervation reflect that the poitive election in migration a related to local farm work i offet by the negative election a related to local nonfarm work, and indicate that the latter electivity i omewhat dominant. In general, the election bia corrected reult how that migrant wage are concave in age. Daily earning increae with age at a decreaing rate; they reach a maximum in the mid-40, and tart decreaing thereafter. We find that there are ignificant gender difference in wage, with men earning about 18% more than women. Education eem to make a poitive difference to earning, although the coefficient on education in the etimation of ubgroup A wa inignificant. With regard to other variable, good health i rewarded in the labor market by higher earning. Young dependent provide an incentive to increae earning. We are urpried to find the coefficient on the dummy of migration experience i negative and ignificant at the 1% level in ubgroup B. Perhap the mot plauible explanation for thi obervation i that repeat migrant pend more time in leiure after they grow into their new environment, and thi negative effect on work hour wa dominant, offetting the expected poitive effect of the higher wage rate. 13 In order to ee the effect of migration experience on the wage rate more clearly, we introduce cro term of migration experience and education in our pecification. Mover who had migration experience in the pat three year are defined a repeat migrant, and thoe who had not moved in the pat three year are defined a new migrant. The reult are contained in Table We find that education doe not affect new migrant earning ignificantly, however ha a poitive effect on repeat migrant earning. It 13 Becaue of data limitation, we cannot control for work hour. 14 Correction term here are obtained from a multinomial logit model with cro term of migration experience and education. The reult of the multinomial logit model are imilar to thoe reported in Table 3-3, and thu we do not preent them here. Thee reult are available upon requet

20 indicate that education enhance migrant earning by increaing their kill through on-the-ob experience over time. It alo provide ome evidence that migration experience ha a poitive effect on the wage rate. More interetingly, we find different return to education in the two ubgroup. Each additional year of chooling ignificantly (at the 5% ignificance level) increae the earning by 3% for pioneer migrant; and the marginal effect of education for migrant from houehold in which ome other member had moved i 1.2% (at the 10% ignificance level). The reult of a one-ided t tet verifie that the return to education i higher for thoe who moved a pioneer than for migrant from houehold in which other member had moved (at the 1% ignificance level). The finding i conitent with the hypothei a we had mentioned that other member migration experience might confer a poitive externality on later migrant (for example, by helping later migrant find a good paid ob) and thu education i more important to thoe without network and thoe who moved a pioneer. Reult for other variable and elective term are very imilar to thoe in Table Concluion In thi paper, we addreed the quetion of elf-election in migration and how it affect migrant earning. Uing data from a rural houehold urvey conducted by the Minitry of Agriculture of China in Sichuan and Anhui province, and utilizing the two-tep election bia correction method developed by Dubin and McFadden (1984) and modified by Bourguignon et al. (2007), we explored the link between employment choice (nonworking, local farm work, local nonfarm work and migratory work) and migration earning. We found that migrant in rural China do not make up a random ample of the population from the village of origin. Our multinomial logit reult ugget that youth, male, the better-educated, and healthy individual are more likely to participate in migratory work. Moreover, earning reult are ubtantively ditorted by election bia a related to unobervable human capital. In particular, imple OLS overetimated the gender difference and the return to chooling, which may lead to fale expectation on migration, for example, providing a diincentive to migrate among female and le-educated peron; or even mileading policy making and policy evaluation uch a public chooling invetment. Unlike mot literature on thi topic, thi tudy not only attribute a election bia in the etimation of earning to the allocation of individual with better or wore unoberved characteritic in migration, but alo link the election bia to the allocation of individual to each other alternative. In the full ample, we find poitive election in migration againt local farm work and nonworking ector and inignificant election againt local nonfarm work. It indicate that the more capable individual (in term of unobervable characteritic) chooe to participate in migration and local nonfarm work, and thoe who are le able are left in the local farm ector or remain out of work. More interetingly, when we ee the reult on the two ubgroup (whether other member had or had not migration experience in the pat three year), we find triking difference between them. For migrant from houehold where other member had not moved (thu making the individual a pioneer), there i no