Migration Duration and Family Economics : Temporary Migration in China and the One Child Policy

Transcription

1 Migration Duration and Family Economics : Temporary Migration in China and the One Child Policy de la Rupelle, Maelys Deng Quheng Abstract Rural-urban migration in China plays an important role in China development and urbanization. Despite important imbalances across the territory, this migration is mostly temporary. This striking feature can be related to institutional constraints, to agriculture seasonality, to unsecure property rights. In this paper, we provide an additional explanation, showing how family planning policies are impacting migrants decision, by constraining ulterior off-farm labor participation decision. No or few siblings means reduced resources in labour power at the household level, which translates into higher constraints on individual decisions. Using the exogenous shift in fertility introduced by the One Child Policy, we compare the elder children born before and after its implementation. We find that individuals subject to this policy are experiencing shorter duration of migration. The Chinese case allows us to document precisely the impact of family size on migration. JEL Classification: D63, P48, O13, I20, J13, O1. Keywords: fertility, migration, household economics, China. This paper has benefited from comments by Marc Gurgand, Terry Sicular, Nancy Qian, Xing Chunbing, Sylvie Démurger, Philippe de Vreyer, Gilles Postel-Vinay, Tatiana Goetghebuer and Imane Chaara. PSE (Paris-Jourdan Sciences Economiques/Paris School of Economics), 48 boulevard Jourdan, Paris, France. delarupelle@pse.ens.fr, phone: (+33)(0) CASS (Chinese Academy of Social Sciences), Yuetan Beixiaojie, Beijing , China. dengqh@cass.org.cn 1

2 1 Introduction Beside having experienced impressive growth and drastic economic changes, China factors markets display two striking macroeconomic features : a high rate of savings, and a high share of temporary migrant workers. Regarding the former one, savings had increased dramatically after the 1980s, skyrocketing to 37% of household income in the mid 1990s. As for the latter, there were 140 million of rural-urban migrant workers in 2008, (according to the National Bureau of Statistics 1 ). Internal migration are not per se surprising - it is a crucial step in countries development and urbanization. What is remarkable in the Chinese case is the nature of these migrations : most of the migrants are temporary migrants. Indeed, in a survey carried on in 2005 by the State Council Research Bureau (see State Council Research Bureau (2006)), only 8.13% of the interviewed migrants declare that they plan a long term stay at their migration destination. Two years later, in 2008, the 5000 migrants surveyed by the RUMiCI survey are only 9.5% to say that, would policy allowed, they would stay forever in cities. These two striking features of the Chinese economy may find some of their origin in China specificities regarding fertility patterns, that have been greatly affected by family planning policies. Banerjee et al. (2010) have shown that the importance of savings rate in urban area can be related to family planning programs. 40 % of the growth of savings rate in the last three decades can be explained by the policies designed to curb population growth. In this paper, we show that the One Child Policy is a determinant of migration temporality in rural China : a reduced number of siblings has an impact on the migration patterns of individuals. The One Child Policy did reduce the total number of children per family in rural area, though it was less effective than in urban areas. We find that the cohorts with less siblings subsequently to family planning policies have systematically experienced shorter duration of migration when they first migrated. Family size has, theoretically, two opposing effects on migration. On the one hand, if migration is used to diversify risk (Stark and Levhari (1982), Stark and Bloom (1985)), one needs to migrate less if one has more siblings. Thus, individuals with fewer siblings are more likely to migrate. On the other hand, if hired labor is not a good substitute for the labor of households members, then fewer siblings may mean that an individual is less likely to migrate. For some households needs, hired labor might be an inferior substitute, like for example the care of elderly parents or children. Tending the farm might as well requires household members labor, especially if their land rights are jeopardized by a long absence, and if the high seasonality of agriculture results in shortages in farm labour 1 See Li (2010) 2

3 supply. Migration itself might be more difficult in the absence of migrating siblings: they could provide information or advices (as shown by Kesztenbaum (2008)). In this paper, we want to estimate the net of these two opposing forces. This is difficult to investigate, both for econometric and practical reasons, linked to the lack of appropriate data and to endogeneity concerns. Family size and migration decision may be affected by a third factor. Parents with a smaller number of kids might adopt different attitudes toward schooling decision. Migrating parents may have a specific fertility, while their experience may help their children to migrate later on. Using a unique survey of migrants conducted in fifteen cities of China in Spring 2008, we deal with these endogeneity issues by using the family planning policies implemented in China at the end of the 1970s, a period where migration were not allowed. Our strategy, building upon Qian (2009), compares migration duration of first borns belonging to cohorts born before and after the introduction of restrictive family planning policies. The understanding of the family planning policies in China requires to take into account their successive steps. A four-year birth spacing law was introduced in The One Child Policy was implemented in 1980s. Combined with the former, the One Child Policy was actually binding for the cohorts born in 1976, as stressed by Qian (2009). We will therefore compare elder children born before and after The results found are the following. Migrants with a lower probability of having siblings have systematically shorter job duration when first migrating. Cohorts bound by the One Child Policy migrate in average 3 months less than previous ones, while the average first job duration is 14 months. To assess the relevance of our interpretation, and to avoid capturing an other effect of the One Child Policy, namely its impact on sex ratio, we also consider a sample with male elders only. We find that the difference goes up to 4 months. We conclude with robustness checks and further comments. The plan of this paper is the following. Section 2 provides elements on the institutional context of rural China, regarding family planning policies, agricultural activities and migration policies. Section 3 describes our data. Section 4 explains our identification strategy. Section 5 displays our results. 2 Institutional Context In this section, we will present briefly some institutional aspects relevant for our study, regarding the family planning policies and the rural-urban migration. 3

4 2.1 Family Planning Policies in China After a period of policies promoting a high natality, the Chinese government started to advocate family planning in the late 1950s. However, no effective birth control measures were implemented before the 1970s. Introduced in 1972, the Wan Xi Shao Policy, meaning Later [age], longer [the spacing of births], fewer [number of children], was encouraging four-year birth spacing, and was recommending three children per rural couple. The One Child Policy (OCP), implemented from 1978 to 1980, was actually announced in Couples complying with the policy were entitled to a one-child pledge certificate and to the economic rewards associated, while unplanned birth were exposing them to fines. Because the One Child Policy was preceded by the four-year birth spacing law, its effect started to appear for cohorts born slightly before the OCP, as shown by Qian (2009). Households who had one child had to wait four years before the next child; if the first birth occurred in the late 1970s, the One Child Policy stopped them from having a second child. Qian (2009) uses the timing and the geography of the relaxation of the OCP to show that its restrictive effect appears in 1976; the cohorts who gained an additional sibling when the One Child Policy was relaxed were the cohorts born in 1976 and after. Given the lack of enforcement and the rural preference for numerous families, the One Child Policy actually bound families at two children rather than one. Kaufman et al. (1989) report how officials eventually regarded a two-child limitation as an adequate policy achievement, while other officials were suggesting that if the government promotes a one-child policy, then people will stop at two; if a two-child policy is promoted, then people will stop at three. The tiny number of households with only one child is further explained by policy changes occurring afterwards. All across China, the family planning policies increased female infanticide, forced abortion and sterilization. After worries raised by the coercive means employed at the local level, as well as the persistent social resistance to birth control, it was decided to loosen these policies in the 1980s. In 1984, a second child was allowed for households facing difficulties or whose first born was a girl, under the condition that a birth spacing was respected, usually of three or four years. A few areas started to relax the One Child Policy slightly before, in 1982 and Many studies have shown that the birth control policies were responsible for much of the drop in fertility rate in rural areas (Shultz and Zeng (1995) and McElroy and Yang (2000)). While few rural households were found to have only one child, the planning 2 Qian (2009) 4

5 policies affected nonetheless their fertility pattern. Statistical evidence To sum up, family planning policies of the 1970s curbed fertility rate in rural China, reducing the potential number of siblings of rural individuals. We do not have census data to show what exactly happened in terms of households fertility, as we do not have detailed information on migrants siblings. Yet, more extensive information on the cohorts born in the 1970s can be found in the rural sample of the RUMiCI project (8000 households). Information on household s 1970s demographic situation is partial, as this survey has been conducted in Data on the number of children per household are biased by deaths and household splittings. Luckily enough, we know the birth rank of the surveyed individuals. If family planning policies have impacted rural households fertility, we should observe variation in the respective share of each birth rank by birth cohort. Births of high order should diminish in cohorts constrained by family planning policies. We should see that the share of children born as the first child increases. In Figure 1, we compute, for each birth cohort, the proportion of first, second, and third and above children. We plot the percentages obtained. We plot a similar graph for the male children in Figure 2. share of individuals of a given birth rank by cohort all - rural sample Percentage in the cohort Birth year of the cohort considered elder child third and above second Figure 1: Share of birth ranks within birth cohort The first remark is in line with what has been previously said regarding policies implementation in rural areas : the One Child Policy did not translate into cohorts of only-child, where all the newborns would be the first child of the households. There are many children who were born as second child in 1980 : the share of second order births remains quite stable until the early 1980s. Yet, we see that in 1980 there is a clear drop in the share of birth of higher order, which becomes almost null. The second remark is that the birth spacing policies implemented before have been 5

6 share of individuals of a given birth rank by cohort men - rural sample Percentage in the cohort Birth year of the cohort considered elder child third and above second Figure 2: Share of birth ranks by cohort - male effective. The share of births of third order and above starts to decrease in 1972, and does even more so four years after, in In Figure 1, we see that although the major part of births in the 1960s were occurring in families having already two or more kids, (they account for more than 40 % of births), there is a clear shift in the mid 1970s, and 1976 is the first year where the share of first borns is higher than the share of births of third or higher order. If we restrict our sample to males, (Figure 2), differences are made clearer. Again, a sharp decrease takes place at the end of the 1970s. share of individuals of a given birth rank by cohort all- migrant sample Percentage in the cohort Birth year of the cohort considered elder child third and above second Figure 3: Share of birth ranks within cohort among rural-urban migrants 6

7 share of individuals of a given birth rank by cohort men- migrant sample Percentage in the cohort Birth year of the cohort considered elder child third and above second Figure 4: Share of birth ranks within cohort among rural-urban male migrants We might be worried that the rural sample fails to account for households whose members are all rural migrants. First, given the size of rural population, it is unlikely that the clear patterns we observe are driven by migration. Still, we provide similar graphics for the different cohorts of the rural-migrant sample. In rural-migrant population, we see as well that individuals born after 1980 are more likely to be the elder child. The share of first born children increases slightly after 1976, as the figures 3 and 4 show. 2.2 Migration, Land and Family Help Various institutions restrict rural-urban migration. Land rights uncertainty affects the mobility of rural outmigrants, as shown by Lohmar (1999), Shi (2004) and de la Rupelle et al. (2009). Migration decision may be affected by parents health, as shown by Giles and Ren (2007). In rural area, help given to origin households is mostly done through labor exchange (Lee and Xiao (1998)). The harvest season is a critical period, where all families may lack labor, and where therefore the help of relatives can be of crucial importance. Not only markets are imperfect; but China institutional system may also increase the role plaid by family size. Through the household registration system, rural households are granted use rights on a piece of land, for which they have to fulfill quota requirements. Household land area depends on household size and household number of laborers; yet, an increased family size, through economies of scale and deeper inclusion in village s labor exchange, releases somehow the constraints attached with tax duties. Restrictions on migration give additional value to household s resources in labor 7

8 power. 3 Data The data used in this paper comes from the RUMiCI (Rural-Urban Migration in China and Indonesia) Project 3. The main originality of this project relies in the method used. The sampling frame was based on a first census conducted among migrant workers at their workplaces. City areas (including suburban areas where factories are located) were divided into blocks of meter. In the randomly selected blocks 4, all the workplaces were surveyed. This census was then used to sample migrants, whose household was also surveyed 5. An important advantage is that we have information on a crucial step of a migrant experience : the very first one. We know the duration of the first job as migrant. Understanding the conditions of the primary migration experience is interesting, and it can help to avoid troubles arising when comparing individuals at various time in their life cycle, who have heterogeneous background as migrants. The experience gained and the contacts established during the first job as migrant may shape a migrant trajectory. The data provides the duration, in months, of the first job as a migrant. We consider it is a good proxy for the duration of the first migration, as staying in cities without employment is more costly than coming back. In the years during which most of the first migration experiences occurred, (second half of the 1990s decade), migration policies were quite restrictive. Even recently, despite considerable loosening of institutional constraints, the massive shutdowns in the aftermath of the financial crisis caused many laid off migrants to come back to their origin village. The main inconvenient is that we do not have the total number of siblings in our data. We know only individual s birth rank. Our identification strategy will have to take into account this shortcoming. The sample we use is of rural-urban migrants, who were surveyed in 15 cities 6. 3 This project, funded by the Australian Research Council, the Australian Agency for International Development (AusAID) and the Ford Foundation, was initiated as a collaboration between the Australian National University and the Beijing Normal University. 4 (accounting for 12 % of each city size) 5 More information on the sampling method are available at the following address : 6 Shanghai, Guangzhou, Shenzhen, Dongguan, Nanjing, Wuxi, Hangzhou, Ningbo, Wuhan, Chongqing, Chengdu, Hefei, Bangbu, Luoyang and Zhenzghou. 8

9 4 Empirical Strategy If family size affects migration, its effect is nonetheless hard to identify properly. Some characteristics may jointly determine fertility decision and ulterior migration of the children. For example, it could be that parents who have a smaller number of kids adopt different attitudes toward schooling decision, that will later have an impact on their situation on the labor market. Interestingly enough, before the 1980s, migration was prohibited, and its control was strictly enforced. The loosening of the structures of collective economy in the 1980s relaxed somehow the constraints affecting migration, without legalizing it. Therefore, the parents who were affected by the family planning programs in the 1970s had very little - if none - experience as migrant. So we should not worry about this endogeneity channel. To deal with other endogeneity sources, we take advantage of China family planning policies that were introduced during the seventies. The introduction of these policies produced an exogenous shock on family size. The number of siblings decreased for cohorts born in the 1970s. We will compare the elder born before and after the implementation of family planning policy. There are three reasons for us to focus on the first born children. First, as birth rank may affect individuals trajectories, and as birth ranks of high order are made rarer by family planning policies, it would be misleading to simply compare people born before and after their implementation : we might capture the effect of a higher share of first born in the population. Second, the number of siblings may have a different impact if they are older or younger. Last, we can consider that the treatment effect of family planning policies is higher for the elders. It is true that all birth orders are affected by a reduced family size. All the individuals born after 1976 should have been the last child of their parents. But the higher the birth rank is, the higher the chance that the family has already reached the desired number of children, and is not constrained by the policy. Moreover, the bigger the family is, the higher the probability is to select a family in an area where the planning policies were poorly implemented. The population of elder children includes all the households who were the most heavily constrained by the planning policies. Let us note l the duration of the first job spell of a migrant. Our identification strategy leads us to focus on the following equation : l = β(elder post p ) + αelder + δpost p + Xγ + ɛ (1) elder is a dummy indicating whether the individuals is the elder or not. post p indi- 9

10 cates that he was born after 1976 and the change in planning policies. X are relevant characteristics, while ɛ is the error term. The interaction term (elder post p ) is our main term of interest, as it indicates the group of individuals who were the most subjects to family planning policies. If the number of siblings eases the family constraints weighing on migration decision, we would expect the coefficient of the interaction term, β to be negative. In such a setting, time-invariant differences across birth ranking will be differenced out by the comparison across time. Then, changes across time which affect elder and non elder similarly will be differenced out by the comparison across birth ranking. As we look at the duration of job spells, it is important to recall that labor market seasonality may have a direct impact on it; regarding the business cycle, different starting period may bring systematically different opportunities regarding job length. For instance, if a given month is a high period of long term workers recruitment, a different starting month will bring systematically a different working duration. To control for urban labor markets seasonality, we introduce job starting month fixed effects. The duration of the first migration spell of a migrant who has started to work in month m can be written as follow: l = β(elder post p ) + αelder + δpost p + ψ m + γx + ɛ (2) ψ m is the month fixed effect. Of course, it only controls for trends affecting all migrants in a similar way. However, it could be that individuals born after 1976 were more likely to face a difficult year, or a change in business cycles trend. They would have had similar patterns than previous cohorts, starting at a similar age, leaving their job after a similar duration. Simply, by entering one year later in the labor market, they were more likely to experience a specific downturn, and to be constrained to leave the city. If it was the case, our estimate would confound the effect of change in macroeconomic situation with the effect of family planning policies. To control for this potential problem, we introduce dummy variables for the year in which migrants left their job, φ t. The last equation we consider is the following : l = β(elder post p ) + αelder + δpost p + ψ m + φ t + γx + ɛ (3) We estimate these equations using a Tobit model, therefore relying on the assumption of errors normality. 10

11 5 Results The rural-urban migrant population surveyed in the RUMiCI project allows us to study the impact of the family planning policies, as the subsamples of migrants born before and after the implementation of family planning policies are both of reasonable size. The migrants sampled in 2008 were 30.5 years old in average. Figure 3 plots the density of the birth year of household heads. 40% of them were born before 1976; the median year of birth is Besides these 5007 migrants, all the individuals staying with the migrant have been surveyed 7. Figure 4 plots the density of the birth year of all household members staying with the head. When excluding the children under the age of 16, the median value of the sample is 1978, while the share of individuals born before 1976 increased at 45 %. To keep cohorts big enough to allow comparisons, but homogeneous enough for the goal of our study, our sample is restricted to the cohorts born after These cohorts were born after the years of natality promotion of the early 1950s, and, more important, after the Great Chinese Famine ( ) which followed the Great Leap Forward and caused a severe drop in the birth rate, while the death rate increased sharply. Moreover, these cohorts started to work after the end of the Cultural Revolution (1976), which makes them even more comparable. In a subsequent set of regression, we also consider an even more homogeneous sample by keeping all individuals who have started to work in a liberalizing economy, by considering cohorts born after Last, we do not consider individuals born after planning policies relaxation. As the first local government to issue permits for a second child did so in 1982 (see Qian (2009)), we keep only cohorts that were born before Complete information on our dependent variable is obviously only available for people who are not in their first migration spell anymore. As the ending date is crucial for our study, we choose to focus on individuals whose current job is not the first job as a migrant. They account for almost two thirds of the cohorts. The average migration duration of their first job as migrant amounts to 14 months. Figure 5 plot the distribution of migration duration, for migrants who were born before and after First, migrants born before after 1976 have, in average, a shorter duration of migration. Second, the distribution is less smooth, and exhibits more seasonality. Given the hight seasonality of agriculture, this could signal that migrants born after 1976 are more likely to come back home for agricultural reasons. Migrants born after 1976 are less likely to have a first job spell ending after 18 to 21 months. A second mode is visible at 24 months, while there is almost no change for migrants born before As our focus is on temporary migration, and to allow more comparability, we do not 7 They amount to 3400 individuals, 2500 of them being above

12 Duration of the first job as migrant months cohorts density: duration_firstjob number of months born after 1976 born before 1976 sample restriction : job as migrant started after 1980 and lasted less than 5 years Figure 5: Density of the duration of migrants first stay in cities consider spells exceeding 5 years, which represent a small share (less than 4%). dependent variable is therefore censored at 1 and 60 months. As written before, the equation we estimate is the following : Our l = β(elder post p ) + αelder + δpost p + ψ + φ + Xγ + ɛ (4) The relevant characteristics, the X, includes characteristics at the individual, household and village level. The individual characteristics that should matter in explaining wage level and employment opportunities are the years of education, the gender, the age when first migrating. Then, we control for household ethnicity, which affects employment and migration opportunities. The main village characteristic we consider is the geographical condition of village, whether it is plain, hilly or mountainous. Village topography, a good proxy for its remoteness, captures difference in educational opportunities, economic dynamism and connection to transportation networks and urban markets. Table 6 in the Appendix displays the descriptive statistics, and provide as well statistics for both groups of elder and non elder. It is difficult to interpret differences across groups, as they also capture differences in age and in family types (as individuals of higher birth ranking are belonging to more numerous families...). Table 1 displays the regression results. From the first regression we run, we see that the interaction term between being the elder child and being born after 1976 has a negative and significant effect. An exogenous decrease in the number of younger siblings has led first born individuals to shorten their migration duration by two months. Given that the average duration of migration among 12

13 Table 1: Migration duration and family size Migration duration and family size Dependent variable : first job duration as migrant (months). Tobit model. (1) (2) (3) (4) (5) first reg with interaction term with land final year dummy head only education 0.387*** 0.393*** 0.347*** 0.252** [0.120] [0.120] [0.129] [0.127] [0.152] male *** *** *** *** ** [0.602] [0.601] [0.637] [0.623] [0.849] age at start *** *** *** *** *** [0.0524] [0.0523] [0.0550] [0.0956] [0.118] elder [0.624] [0.822] [0.862] [0.842] [1.017] born after *** *** *** *** *** [0.666] [0.789] [0.828] [1.121] [1.357] minority [2.109] [2.108] [2.257] [2.201] [2.537] elder X born after * ** ** ** [1.252] [1.313] [1.279] [1.537] log hh land * ** [0.510] [0.499] [0.593] hometown geography yes yes yes yes yes starting month dummies yes yes yes yes yes final year dummies no no no yes yes Constant 17.15*** 16.77*** 17.88*** 46.45*** 45.15*** [2.103] [2.110] [2.234] [12.37] [12.74] Observations Pseudo R Standard errors in brackets * p < 0.10, ** p < 0.05, *** p < 0.01 Sample restricted to cohorts whose first job initiated after 1980 and lasted less than 5 years 13

14 the sample considered is 14 months, it is a non negligible effect. The other variables have non surprising effects. Education plays a positive role on duration, though a small one, implying that more educated migrants are also more stable in their first working experience as a migrant. The dummy male has a negative and significant effect. We believe that male migrants shorten their migration to participate in agricultural work, more than female migrants. If household involvement in agricultural work matters, we should find a sharper effect when excluding the migrants whose households do not use any farmland, and may therefore have more flexible work schedules. To test this hypothesis, we drop the 200 migrants coming from a household who does not hold any land, and we control for the logarithm of land per person in the family. 8 The regression is presented in the third column of Table 1. In this regression, the coefficient of the term of interest is more significant, which is consistent with our interpretation. As mentioned in the previous section, such a setting fails to control properly for business cycles or year specific shocks. It would be worrisome if elders, having migrating earlier, were more likely to face this special year, either at their destination area or at their home village, that would have require them to go back home. Our specification fails to capture shocks affecting differently elder and younger siblings both born after Therefore, we add dummies controlling for the year in which the migrant interrupted his job. The regression is presented in the column four of Table 1. Our results remain unchanged. In all regressions, we have considered all the surveyed individuals : the migrants selected through the sampling process, and the household members living with them. They might be migrants kids or spouse. In the regression displayed in the fifth column of Table 1, we restrict our sample to the initial migrants sample. The column is labeled head only 9. Our results hold : the coefficient of the interaction term is still negative and significant at the 5% level. 8 Though the mean of land endowment in origin household is at around two mu per person (One mu is equal to 666,67 m 2 (25,8m 25,8m)), around 40 migrants of our sample rely on 10 mu and more per person. We take the logarithm so that our findings will not be driven by families relying on a big amount of land. 9 It is important to note that this term does not refer to intra-household decision making process but to the survey sampling frame. 14

15 5.1 Robustness checks Alternative samples Sector specific shocks When interpreting these results, one may want to know whether this shortened duration is reflecting macroeconomic shocks or if it is the result of a choice. Luckily enough, we know the condition under which the job ended, namely whether the job was left voluntarily, or not. There is no reason to believe that a firm would prefer to fire employees without siblings rather than employees with a numerous family. Although mechanisms are not very clear, we could imagine that if elder versus non elder are employed in very different types of sector, and if a specific sectoral shock decreases considerably job duration for the youngest cohorts, it would impact the coefficient of the interaction term. One way to check that we are not actually capturing differences in sectoral shocks is to drop the 300 migrants in our regression sample who did not left voluntarily their job. If the effect disappears, this will put into question our interpretation. If the effect persists, though it will not rule out alternative explanation, at least it will be consistent with our explanation. When we exclude them, we see that our results are holding, as shown in Table 2. The coefficient of interaction term is still significant at the 5% level. Gender A more serious concern relates to the main characteristic that changed for elder who were born before and after the implementation of the One Child Policy : the gender. The One Child Policy had an impact on the sex ratio, it increased the practice of female infanticide or selective abortion. Our result may capture the fact that babies born after 1976 were mostly males. We are then just capturing a change in gender composition within cohort. However, in 1976, only the 4-year birth spacing policies was enforced - the One Child Policy was to be announced and implemented four years later. The households did not know that their baby was meant to be the only one, so sex selection was not as dramatic as in the early 1980s. It does not mean that sex selection did not occur, but there is no reason to believe that it suddenly increased between 1975 and Last, if birth spacing has given rise to sex selection, then it would have started in 1972, when the policy was implemented, not in Moreover, in the 1980s, planning policies were partially relaxed in rural areas, allowing families with only one girl to have a second child. As a result, first born women born after 1976 have usually gained a sibling afterwards; Qian (2009) confirms it. The treatment effect captured by our term of interest is therefore weaker on female elder. So not only we hope to obtain a significant coefficient for our interacted term, but we hope also it will have a bigger magnitude. 15

16 Table 2: Migration duration and family size when job was left voluntarily Migration duration and family size - Job was left voluntarily Dependent variable : first job duration as migrant (months). Tobit model. (1) (2) (3) (4) with land final year dummies male only male head only education 0.273** [0.135] [0.134] [0.173] [0.183] male *** ** [0.664] [0.651] age at start *** *** *** *** [0.0585] [0.101] [0.128] [0.141] elder [0.912] [0.890] [1.124] [1.221] born after *** *** *** *** [0.863] [1.173] [1.499] [1.621] elder X born after ** ** ** ** [1.368] [1.331] [1.668] [1.819] minority [2.303] [2.241] [2.765] [2.880] log hh land * ** [0.531] [0.519] [0.637] [0.691] hometown geography yes yes yes yes starting month dummies yes yes yes yes final year dummies no yes yes yes Constant 18.77*** *** [2.340] [11.90] [11.78] [6.609] Observations Pseudo R Standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 Sample restricted to cohorts whose first job initiated after 1980; lasted less than 5 years; was left voluntarily 16

17 We run regressions restricted to male migrants, that we present in the column 3 and 4 of Table 2. Our results remain unchanged, and the effect is stronger. Agriculture The principal concern we have is related to the accuracy of our interpretation. We have argued that agricultural work was of crucial importance; we run additional regressions to check that this is actually the case. First, migration duration will be shortened in order to help the household during the peak agricultural season only if the family is actually farming its land. Households who have given away their land to neighbors, or who are subcontracting it should be far less demanding than households working on their plot. Their might be other reasons constraining migrants decision and leading individuals with less siblings to come back earlier; yet, as we suspect that needs related to agricultural work are of primary importance for many rural households, the impact of the interacted term should be less significant for families who do not have these needs. We run two regressions, restricting our sample first to migrants whose land is farmed by family, second to migrants whose land has been subcontracted or given away. The results, displayed in Table 3, are consistent with our interpretation. In the sample where land is subcontracted or given away, standard errors are much higher than in the sample where land is farmed by family. The coefficient does not changed its value, but it is not significant anymore. One of the reason is that our sample size is smaller. However, other regressions with small sample size keep the coefficient significance (see below). Similarly, if our reasoning is correct, then we should find a stronger effect for the migrants who were likely to have participated in agricultural work before migrating. If their family lacks labor power, then the migrants were probably already involved in farming activities when they were living in their home village. They would not have then left their home during the harvest season, nor slightly before. We remove the individuals who have left home between June and November, a period of intensive agriculture activity. These individuals were not constrained by household agricultural requirements the year during which they left. The coefficient of the interaction term conveys a stronger impact; it is even more significant. Restricting further our sample by considering only men gives further support. (See column 3 and 4 of Table 3.) 17

18 Table 3: Farming work and migration Robustness check: farming work and migration Dependent variable : first job duration as migrant (months). Tobit model. who does the farming work? migrant left home before June or after November Land is : did not leave during peak season (1) (2) (3) (4) VARIABLES farmed by family subcontracted, given away all men only education [0.160] [0.247] [0.148] [0.192] male ** [0.750] [1.287] [0.731] age at start *** *** *** *** [0.140] [0.177] [0.113] [0.146] elder ** [1.108] [1.649] [1.001] [1.274] born after *** *** *** *** [1.374] [2.480] [1.317] [1.700] elder X born after ** *** *** [1.509] [2.919] [1.510] [1.901] minority * [2.821] [4.401] [2.750] [3.660] log hh land * [0.606] [1.011] [0.587] [0.731] hometown geography yes yes yes yes starting month yes yes yes yes final year yes yes yes yes Constant 44.89*** 52.01*** [12.19] [9.118] [11.61] [11.49] Observations Pseudo R Standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 Sample restricted to cohorts whose first job initiated after 1980; lasted less than 5 years Additional sample restriction is indicated in column header 18

19 5.1.2 Placebo regressions Yet, a question remains. Our results seem to be related to family needs, notably in terms of agricultural work. Do we actually capture the effect of family planning policies, or do we confound it with broad changes that occurred after the opening of the economy? It could be that individuals born after 1976 were more likely to face a difficult year, or a change in business cycles trend. Younger and elder cohorts could also have had different experiences, strongly affecting their migration opportunities. For example, the children born in 1976 and after were more likely to be educated in a different context, and to enter primary schools in the 1980s. After the reforms, educational quality changed: rural enrollment rates decreased substantially, while many rural schools of low quality were shut down (Hannum et al. (2007)). This could be worrisome if it affected elder children in a specific way. If the elder child was the only one to be educated, then a change in education quality would affect elder and non elder differently. However, the schools shut downs mainly affected junior secondary schools (Hannum et al. (2007)). As recalled by Qian (2009), the degradation of education quality was important for middle-schools and highschools, while the change in primary education was negligible. The sample was chosen in order to be able to keep comparable cohorts. However, the pre-1970 cohorts may have started to work before liberalization process initiates; to further check that our effect still holds for cohorts more similar, we run a few regressions on a restricted sample, only taking into account the cohorts. All these cohorts may have had comparable middle-school education. Results are holding, as shown in column one and two of Table 4. Our effect is therefore not explained by differences between 1960s and 1970s cohorts, and holds when doing our investigation within the 1970s cohorts. To deal with similar concerns on specific time trends affecting elder, we run a set of Tobit regressions, keeping everything identical but the year used to compare elder cohorts across two groups. We expect that the coefficient of the interaction term would not be significant for these placebo years. First, we consider the full sample. The effect previously obtained is driven by the comparison across younger and older cohorts, so we should still find an effect when splitting the sample for other years, as there will still be a majority of unconstrained people in the cohorts born before the placebo year. So in a second step, we will drop all migrants born after The coefficient obtained should be very different. 19

20 Table 4: Migration duration and family size cohorts Migration duration and family size cohorts Dependent variable : first job duration as migrant (months). Tobit model. (1) (2) (3) final year dummies head only male only education [0.145] [0.173] [0.188] male ** ** [0.687] [0.917] age at start *** *** *** [0.183] [0.218] [0.227] elder [1.078] [1.284] [1.330] born after *** *** *** [1.339] [1.597] [1.658] elder X born after ** ** ** [1.413] [1.682] [1.758] minority [2.304] [2.639] [2.880] log hh land [0.546] [0.645] [0.662] hometown geography yes yes yes starting month dummies yes yes yes final year dummies yes yes yes Constant 45.59*** 32.28*** 29.88*** [11.62] [11.39] [8.571] Observations Pseudo R Standard errors in brackets * p < 0.10, ** p < 0.05, *** p < 0.01 Sample restricted to cohorts whose first job initiated after 1980 and lasted less than 5 years 20

21 In the figure 6, we plot the coefficients of the interaction term (elder)*(born after year Y) obtained for different years Y. The only year exhibiting a coefficient significant at the 5% level is In the figure 7, we do the same, but restrict our sample to individuals born before Thus we drop the individuals affected by the One Child Policy. We do not obtain significant coefficients anymore. Even more interestingly, the coefficient value is positive, and it seems to follow an increasing trend. Then, we plot similar figures for the population of male only, which should be more affected by family planning policies. In figure 8, two years exhibit significant coefficients : 1976 and The result is interesting, as in 1972 the four-year birth spacing law was implemented. This shows that our effect is related to family planning policies. When we look at the results obtained on the sample of male migrants born before 1976, displayed in figure 9, the comparison is very striking. Figure 6: Coefficient of the interaction term between being an elder and belonging to cohorts born after the year x. Along with the x axis, the year used as a placebo, splitting the two samples, varies. 21

22 Figure 7: Coefficient of the interaction term between being an elder and belonging to cohorts born after the year x. Along with the x axis, the year used as a placebo, splitting the two samples, varies. Sample restricted to migrants born before Figure 8: Coefficient of the interaction term between being an elder and belonging to cohorts born after the year x. Along with the x axis, the year used as a placebo, splitting the two samples, varies. Male only 22

23 Figure 9: Coefficient of the interaction term between being an elder and belonging to cohorts born after the year x. Along with the x axis, the year used as a placebo, splitting the two samples, varies. Male born before 1976 only. Migration location Finally, we may be concerned by the fact that job durations of migrants are shorter because they prefer to leave their place of work in order to search for jobs in other cities. Indeed, in Bhattacharya (1990) s theoretical framework, migration temporariness is explained by a search process in various locations. Data provides us with the number of cities/towns where individuals have ever migrated for work purpose. We run regressions with this variable as dependent variable, both with OLS and Tobit model. Our variable of interest has no impact on the number of cities that a person has ever migrated to. Therefore, shorter migration duration can not be related to an increased number of migration locations. 5.2 Comment on sample selection As mentioned before, we have been doing our study on migrants whose current job is not their first job as a migrant. This implies a sample selection along two dimensions. First, among the urban migrants, we do not consider permanent migrants. As said before, a definitive settling in an urban area is difficult and therefore often out of reach for most rural outmigrants. Nevertheless, migrants who never came back home will not be included in our sample. Second, we do not account for temporary migrants who may migrate only once during their working life, and never come back in a city afterwards. It can be the case if the first experience as a migrant was a bad experience (if for example they did not get paid at the end, as it could happen for the first generation of migrants, whose working conditions were especially harsh) or after a dramatic change in their household 23

24 situation at home. Regarding the former concern, a first way to answer it is to include in our sample the migrants we had dropped and run similar regressions. The dependent variable becomes much noisier, especially as the survey was not done at the same month for all. Most of migrant households were interviewed in May or April (56% of the households were surveyed in May; 32% in April), still, 11% of them were surveyed in March, June, and even August for a handful of them. To account for these issues, duration models would be more appropriate. The use of such models is planned for the next step of our research project. The Tobit model already allows us to obtain promising results. The main change made from previous regression is that an other dummy is added to the set of dummies indicating the year in which the migration spell we terminated. It equals to one when the migration spell ended after the survey was done. The last table, Table 5, shows that the term of interest is still negative and significant. Regarding the latter concern, one solution would consist in extending our analysis to the rural sample of the RUMiCI survey. The short term migrants who had only one experience as a migrant could, hopefully, be observed in this sample. Last, an other interesting direction of research is to provide empirical elements showing the relevance of the first migration spell in migrants trajectories. 24

25 Table 5: Sample including individuals whose current job is their first job Migration duration and family size - Including unfinished migration spells Dependent variable : first job duration as migrant (months). Tobit model. (1) (2) (3) (4) Final year dummy With land Male Male heads education 0.232** 0.204* 0.324** [0.115] [0.123] [0.159] [0.168] male [0.563] [0.599] age at start *** *** *** *** [0.0831] [0.0877] [0.115] [0.124] elder [0.761] [0.802] [1.022] [1.099] born after *** *** *** *** [1.031] [1.086] [1.408] [1.499] elder X born after ** ** *** ** [1.173] [1.237] [1.576] [1.700] minority [2.016] [2.191] [2.646] [2.692] log hh land ** [0.466] [0.577] [0.614] origin village geography yes yes yes yes starting month dummies yes yes yes yes final year dummies yes yes yes yes Constant *** [13.02] [12.93] [12.64] [6.223] Observations Pseudo R Standard errors in brackets * p < 0.10, ** p < 0.05, *** p < 0.01 Sample restricted to cohorts whose first job initiated after

26 6 Appendix Table 6: Descriptive statistics and birth rank Individuals All Are Are not the elder the elder Variables Mean St. Dev. Mean St. Dev. Mean St. Dev. Individual characteristics education (years) male elder born after Household characteristics land (mu) from an ethnic minority Village characteristics plain hills mountains First migrant job duration (months) age at start starting month year of job ending job left voluntarily Observations Sample restricted to cohorts whose first job initiated after 1980 and lasted less than 5 years References Banerjee, A., Meng, X., Qian, N., September The life cycle model and household savings : Micro evidence from urban china. Tech. rep., MIT and Yale University. Bhattacharya, G., October Migration under uncertainty about quality of locations. Journal of Economic Dynamics and Control 14 (3-4), de la Rupelle, M., Deng, Q., Li, S., Vendryes, T., Land rights insecurity and temporary migration in rural china. IZA Discussion Paper 4668, Institute for the Study of Labor (IZA). Giles, J., Ren, M., May Elderly parent health and the migration decision of adult children: Evidence from rural china. Demography 44 (2), Hannum, E., Behrman, J., Wang, M., Liu, J., Education in the reform era. In: Brandt, L., Rawski, T. (Eds.), China s Great Economic Transformation. Cambridge University Press.