The Impact of Parental Migration on Cognitive Ability
Scores of China’s Left Behind Children
Author: Livia Remeijers 1st supervisor: Student Number: 10217002 Prof. dr. M.P. Pradhan
MSc Economics 2nd supervisor:
Field: Development Economics Prof. dr. E.J.S. Plug
Academic year 2015/2016
1 Statement of Originality
This document is written by Student Livia Remeijers who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
2 Table of Contents Abstract ... 3 Acknowledgement ... 4 1. Introduction ... 5 2. Theoretical Background ... 7
3. Data & Analytical Sample ... 10
4. Variables ... 11 5. Methodology ... 14 5.1. Statistical Model ... 14 5.2. Assumptions ... 16 6. Results ... 18 6.1. Descriptive Statistics ... 18
6.2. Propensity Score Model ... 18
6.3. Weight and Balance Diagnostics ... 19
6.4. Impact of Parental Migration ... 24
7. Results by Gender ... 27
7.1. Preliminary Steps ... 27
7.2. Impact of Parental Migration by Gender ... 28
8. Discussion ... 29
9. Conclusion ... 32
References ... 33
3 Abstract
China’s large rural to urban migration flow resulted in millions of left-behind children. Using data of two waves from the China Family Panel Studies, this thesis examines the impact of parental migration on cognitive ability scores of children aged 12-18 left behind in rural China. Inverse probability of treatment weighting using propensity scores is applied to account for systematic differences in observables between left-behind and non-migrant children. Parental migration has a significant negative effect on children’s cognitive ability scores. In addition, the socio-economic effects underlying the overall effect of parental migration are illustrated by an increase in per capita family income and a decrease in the maximum education level in the household. The results suggest that the negative social effect outweighs the positive economic effect of parental migration. While this also holds for left-behind boys, left-behind girls seem to be unaffected by migration of at least one parent. The foremost explanations for the gender difference are father absence, a shift in household decision-making authority, and a stronger impact of remittances on educational investments in girls.
4 Acknowledgement
First of all, I would like to express my gratitude to my supervisor Prof. dr. Pradhan for introducing me to the topic of China’s left-behind children, for his guidance, his interesting comments and valuable suggestions throughout my master thesis process. Besides, I would like to thank him for his instructive lectures during my master degree. My sincere thanks also goes to my other professors from the University of Amsterdam whose lectures I have followed on Development Economics. Besides, I would like to thank Prof. dr. Plug for reading my thesis as second supervisor. Furthermore, I am incredibly grateful for my parents’ love and support. In particularly I thank my father for being so close to me towards the end of the thesis project. Last but not least, I would like to thank Caspar Wanders for showing his enthusiasm for my topic, for reading parts of my work and for his encouragement.
5 1. Introduction
This thesis seeks to explore the causal effect of parental migration on cognitive ability scores of left-behind children in China. Left-behind children are defined as rural children living separately from at least one parent due to parental migration. Although migration is a worldwide demographic process, the largest rural to urban migration in human history has been witnessed in China. Since the economic reforms (1978>) and the opening up of its market, millions of people have moved from the countryside to the cities in search of work. Thus economically driven migration has contributed to an extensive increase in urban population from 19% in 1980 to 56% in 2015 (World Bank, 2016). The National Bureau of Statistics of China reported a total number of 169 million rural migrants working in urban regions in 2015 (National Bureau of Statistics, 2016). It is continuously observed that migrant parents leave their children behind, to be looked after by a single parent or the grandparents. The foremost reasons for not bringing their children along are high migration costs and the restrictions imposed by the so-called hukou system. Hukou is China’s household registration system, which prevents rural registered children from using urban public health and education services (Liang, Guo & Duan, 2008). As a result, about 100 million children, one third of China’s total underage population, are left-behind by at least one parent in 2016 (All-China Women’s Federation, 2016).
Despite a growing literature on the left-behind children phenomenon, there is no consensus regarding the gross effect of parental migration on left-behind children. The unknown total effect is caused by two factors of a socio-economic nature. On the one hand, migrated parents generate higher earnings, which can reach left-behind families through remittances. These extra economic resources can lead to an increase in spending on health and education, hence promoting child development and improving children’s educational prospects (Nguyen, 2016; Adams & Page, 2005). On the other hand, following Lu (2012) this thesis assumes that alternative caregivers, such as grandparents, are less educated than the parents. Therefore, parental absence due to migration can leave left-behind children in a poorer domestic educational environment, which in turn can negatively impact children’s educational performance (Zhang, Behrman, Fan, Wei, & Zhang, 2014).
6 Drawing on 2010 baseline data and 2012 follow up data from the China Family Panel Studies (CFPS), a nationally representative longitudinal survey, this thesis answers the following research question: What is the impact of parental migration on cognitive ability scores of China’s left-behind children? And how might this vary by gender? To answer this question, I condition on children living in non-migrant families at baseline in 2010, and compare cognitive ability scores of children left-behind by 2012 with scores of non-migrant children.1 This thesis contributes to the literature in three important ways. First, applying inverse probability of treatment weighting using propensity scores, accounts for systematic differences in observables between treated (i.e. left-behind) and untreated (i.e. accompanied) children. To my knowledge, this study is the first to use this method in this particular context. Second, I aim to disentangle the socio-economic mechanisms underlying the overall effect of parental migration on cognitive ability of left-behind children. Third, besides performing the analysis for all left-behind children, I stratify the analysis by gender, which enables me to examine whether there is heterogeneity in the effect of migration on girls relative to boys.
This study finds an overall negative effect of parental migration on cognitive ability scores of left-behind children. As expected, per capita family income increases, and the maximum education level in the household decreases as a result of parental migration. This thesis suggests that for the full sample the negative social effect outweighs the positive economic effect of parental migration. While this also holds for left-behind boys, left-behind girls seem to be unaffected by migration of at least one parent.
The remainder of the thesis is organized as follows: part two reviews the literature by describing the channels through which parental migration can affect left-behind children, and reports findings of previous empirical studies. Followed by a description of the data and the analytical sample in part three. The variables included in the propensity score model are explained prior to the statistical model and assumptions. The results of this thesis are presented in part six and seven, and are discussed in part eight. Part nine concludes.
1Non-migrant children are rural children whose both parents live at home. I will also refer to
7 2. Theoretical Background
Parental migration is a household decision rather than an individual one. The main reason for parents to migrate is to improve children’s life opportunities by maximizing household economic welfare (Lu, 2014). The China Youth Research Center (2006) reported that more than 90 percent of rural Chinese migrants with children send remittances back home. Previous studies have shown that these extra economic resources increase household income, and can improve children’s development and education (Edwards & Ureta, 2003; Calero, Bedi & Sparrow, 2009). Therefore, I expect left-behind children to benefit from higher family income and increased educational expenditures.
Yet, parental migration comes at a social cost. Parental absence leads to the lack of an authoritarian figure, less support and less supervision that are fundamental to child development and educational achievement (Parreñas, 2005). Furthermore, in the context of China, contact between left-behind children and migrant parents is generally infrequent. It is common for rural migrants to visit home only once a year – for the Lunar New Year. McLanahan and Sandefur (1994) argue that parental absence significantly affects education and psychological well being of left-behind children. However, in case alternative caregivers are good substitutes of migrant parents, they could mitigate the negative effect. In China, grandparents commonly take over the care of their grandchildren when parents migrate. Considering Lu’s (2012) statement that rural grandparents lack education, it could be that the maximum education level in the household decreases when parents migrate. A lower domestic educational environment is expected to disadvantage left-behind children.
Furthermore, the child development literature suggests that girls and boys are affected differently by parental absence. For instance, Santrock (1972) and Thomas (1994) both argue that sons are more negatively affected by father absence than daughters. Worldwide, it is generally the father who migrates rather than the mother, if not both. Hence, I expect boys to be disadvantaged by parental migration. Nevertheless, there is a strong male preference in rural China, which ensures that financial resources are first distributed towards sons; any extra resources are allocated to daughters. Therefore, boys often receive more schooling than girls (Parish & Willis, 1993). Extra
8 financial resources, as a result of parental migration, are therefore expected to benefit daughters more than sons.
The empirical literature shows that the evidence on the causal effect of parental migration on left-behind children is mixed. First of all, the paper by Xu & Xie (2015) comes closest to my study. Drawing on 2010 baseline data from the CFPS, they apply propensity score matching in estimating the effect of migration on Chinese rural origin children’s well-being. Their results suggest that left-behind children and accompanied children do not significantly differ in terms of cognitive ability, health and subjective well-being. On the contrary, rural migrant children seem to benefit from moving to the city in every aspect relative to rural children. Yet, rural migrant children are worse off in subjective well-being when compared to urban children. An important limitation to their study however, is the data use of a single wave. Therefore, their propensity score model is restricted to accurate pre-migration factors.
Nguyen (2016) also estimates the effect of parental migration on health and cognitive ability of left-behind children in Vietnam, India, Ethiopia and Peru using individual fixed-effects. While no effect is found on left-behind children in Ethiopia, children’s health seems to be negatively influenced by parental migration in Vietnam, India and Peru. In addition, the author finds a negative link between parental migration and cognitive ability of left-behind children in Vietnam and India. He also finds evidence for heterogeneity in the effect of gender of the migrant parent. In Vietnam maternal migration seems to have a stronger negative effect on left-behind children, while in India paternal migration is more harmful. Finally, Nguyen suggests larger negative effects of long-term parental migration.
Moreover, several studies find a positive relation between children’s education and migration of parents. Intemann & Katz (2014) find evidence that left-behind children complete more school years than accompanied children in El Salvador. Furthermore, Antman (2012) suggests that Mexican girls benefit from paternal migration to the US, but not from paternal domestic migration in terms of school years completed. However, other studies find adverse effects of parental migration on children’s schooling. McKenzie & Rapoport (2011) report a negative effect on school attendance and attainment for left-behind children in Mexico. Moreover, Lu (2014) finds that parental migration more
9 adversely affects children in Mexico, than children in relatively more resource-constrained Indonesia. In addition, results by Zhang et al. (2014) suggest that test-scores of left-behind children in China are negatively affected when both parents migrate, but not when a single parent migrates. In summary, the empirical evidence is context dependent and suggests that no general answer exists to the question of the impact of parental migration on left-behind children.
10 3. Data & Analytical Sample
This thesis uses data from the China Family Panel Studies (CFPS), a nationally representative longitudinal survey led by the Peking University. The survey is conducted in 25 provinces, and covers topics such as family dynamics, migration, economic activities, education and health – in urban and rural regions. According to Xu and Xie (2015) CFPS is the first social science survey in China to measure cognitive ability in a nationally representative survey. So far, the 2010 baseline survey and the 2012 follow up survey have been made public. The baseline survey successfully interviewed 14,960 households and 42,590 individuals, of which 33,600 adults and 8,990 children. The response rate is 84.1% in 2010 and the successful-tracking rate is 80.6% in 2012 (Xie & Hu, 2014).
For the purpose of this study, the constructed dataset includes information on parental migration, family structure, family’s socio-economic status, and education. The analytical sample consists of rural children aged 12 to 18 in 2012, as various observables (e.g. test scores) are only available from the age of 10 at baseline in 2010. The sample is limited to rural children with parents living at home or parents being absent because of migration. Other reasons of parental separation such as death and divorce are not considered. Furthermore, I decided to condition on children living in non-migrant families at baseline. Thus, children whose parent(s) migrated before 2010 are not included in the analysis. This approach allows me to estimate the immediate average effect of parental migration on cognitive ability scores of children left-behind by 2012. Unlike with cross-sectional data, longitudinal data enables me to use information from before and after the event of migration, and so investigate a dynamic relationship. Nevertheless, using longitudinal data also comes at a cost, that of attrition. My study suffers from 19.4% attrition between the 2010 survey and the 2012 follow up survey, which is in line with the successful-tracking rate of 80.6% in 2012. I make use of a balanced panel over two time periods (i.e. 2010 and 2012), involving 1,215 rural Chinese children from 1,022 families, of whom 618 are girls and 597 boys, aged 12 to 18 in 2012.
11 4. Variables
The key dependent variable, children’s cognitive ability, is measured through a word recall test (scale: 0-10) and a number series test (scale: 0-15) in 2012. In the word recall test children had to remember as many words as possible of a set of 10 words. The amount of words remembered right determined their word recall test score. The test was designed in a way that it was difficult to remember all words. In the number series test, children were asked to fill in the blank in a series of numbers. The amount of right answers determined their number series test score. To make the two test scores comparable in terms of scale, the number series test scores are re-scaled to the scale of 0-10, by multiplying all its values by 2/3. The estimation model will be applied to both measures of cognitive ability. Other dependent variables, representing the economic and social effects of parental migration are, log per capita family income, log family expenditure on education for child i, and the maximum education level in the household (on a scale of 1: illiterate/semi-illiterate to 7: master’s degree). It is important to note that accurate data on remittances is missing for the year 2012. However, this omission will be addressed through a significant increase in per capita family income as will be evident from my results presented in part 6.4. In addition, reliable data on children’s school years completed is missing; therefore the dependent variables school attendance and children’s current school level are added as alternative measures of educational investment.
Furthermore, the treatment variable parental migration status in 2012 is derived from whether child i has both parents living at home or is separated from at least one parent due to migration, conditional on that child i was living in a non-migrant family at baseline in 2010. Parental migration status by 2012 is reported in Table 1. Of all children in the sample, 10.8% is left-behind by at least one parent, while the remaining 89.2% has both parents living at home. From the table below it can be observed, that the majority of the left-behind children in the sample live in a father-only migrant household.
Despite a lack of consensus on which variables to include in the propensity score model, various studies argue in favor of inclusion of all factors that influence the outcome (i.e. the potential confounders) and all factors that are associated with both treatment exposure and the outcome (i.e. the true confounders).
12 Table 1.
Parental migration status by 2012, conditional on children living in non-migrant families at baseline in 2010 (ages 12-18 in 2012; N=1,215)
Parental migration status N Percentage of sample (%)
Non-migrant 1,084 89.2
Migrant 131 10.8
Single parent migrant 112 9.2
Father-only migrant 88 7.2
Mother-only migrant 24 2.0
Both parents migrant 19 1.6
Source. Author’s calculation
There is evidence that these two sets of variables improve precision of the estimates, while limiting additional bias (Brookhart et al., 2006; Austin, Grootendorst, & Anderson, 2007). Besides, all included covariates should be pre-migration measures or variables that are not affected by the event of migration (e.g. age and gender) (Austin, 2011b). The main advantage of conditioning on children in non-migrant families at baseline is the ability to balance the data on pre-migration variables.
Drawn on published literature relevant individual, family and community level characteristics are incorporated in the treatment model. Children’s demographic characteristics are controlled for by including age and gender. Variables as family size and the number of generations in the family describe the family structure. Three generations within a family implies that at least one grandparent is still alive. Hence, the presence of a third generation might indicate the availability of alternative caregivers, which could ease the decision of parents to migrate, hence affecting treatment exposure. On the other hand, the presence of grandparents inclines to benefit child development by offering supplementary resources and intergenerational support (Falbo, 1991). Therefore, extended kin might have a positive effect on cognitive ability through improved child development. Other true confounders, that approximate family’s socio-economic status, are parent’s years of completed schooling and the maximum education level in the household (measured on a scale of 1: illiterate/semi-illiterate to 7: master’s degree). Furthermore, per capita family income at baseline is an essential ingredient of the propensity score model, as higher urban incomes are the foremost motivation for rural
13 parents to migrate. Higher incomes that reach the left-behind through remittances can have a positive effect on children’s cognitive development as long as they are invested in them (McKenzie & Rapoport, 2011). Xu and Xie (2015) however, exclude this variable from their propensity score model, because they do not have baseline data, hence all observables (i.e. covariates, treatment, and outcomes) are measured in the same year. Therefore, their treatment model is restricted to time invariant variables or variables that are not influenced by the event of migration. An advantage of my study is the use of 2010 baseline data for all baseline covariates, and 2012 follow up data for the treatment and outcome variables. As a result, per capita family income at baseline is included in the propensity score model, while increasing precision and reducing bias of the estimates. Additionally, broader social and environmental variables, such as the travel time to the provincial capital and the percentage of agricultural labor force at the community level are included.
Moreover, the potential confounders are represented by family expenditure on education for child i, and children’s educational characteristics e.g. school attendance and the current school level (measured on a scale of 1: illiterate/semi-illiterate to 7: master’s degree). Finally, pre-migration mathematic grades, Chinese grades, and cognitive ability scores are added to the propensity score model (see Table A1 for the descriptive statistics of the baseline covariates). It is important to note that cognitive ability is measured differently at baseline in 2010 than in 2012. The set of tests used at baseline contains literacy and mathematic questions and is indicated by the variable names “word test” and “math test”. The set of cognitive ability tests used in the follow up survey in 2012 consists of memory and number series questions indicated by different variable names: “word recall test” and “number series test”. The CFPS rotates the two sets of cognitive ability tests by wave (Xie & Hu, 2014). Although, the 2010 and 2012 cognitive ability tests are not identical, the cognitive ability test scores at baseline are a good predictor of the 2012 cognitive ability test scores. The before mentioned variables: per capita family income, family expenditure on education for child i, children’s current school level and the maximum education level in the household are baseline outcome variables.
14 5. Methodology
5.1. Statistical Model
Unlike in randomized experiments, treatment in observational studies is nonrandom, such that the average treatment effect cannot be estimated by directly comparing the outcome means between treated and untreated subjects. In the context of this study, a child is said to receive treatment if at least one parent migrated; thus the treatment is dependent on a choice variable. Therefore, left behind children and accompanied children are likely to differ in terms of their characteristics. To adjust for systematic differences between treated and untreated children this thesis follows Rosenbaum (1987) in applying inverse probability of treatment weighting (IPTW) using propensity scores. This method creates a synthetic sample in which the distribution of observed characteristics is similar across treated and untreated children. Consequently, an unbiased estimate of the average treatment effect on the treated (ATT) can be computed; that is, the average effect of parental migration on cognitive ability scores of children whose parent(s) migrated. The related causal question here is: what would the cognitive ability scores of child i be, if her parent(s) migrated compared to if both parents lived at home? Applying the Rubin Causal Model the ATT is defined as (Rubin, 1974):
[ | ] [ | ] [ ] (1)
In equation (1) is the outcome for child i when treated (i.e. left-behind), and is the outcome for the same child when untreated (i.e. accompanied). implies that child i is treated and that child i is untreated. Since, only one outcome per child is observed, a legitimate counterfactual has to be estimated. This can be achieved by computing the propensity score through a logistic regression:
(2)
In equation (2) is the predicted probability that child i receives treatment in 2012 given the observed 2010 baseline characteristics X. As a result, left-behind children and
15 accompanied children with the same propensity scores can be compared, as the distribution of the propensity scores is similar. A next step is to weight the entire sample using the predicted propensity scores to ensure that treatment exposure is independent of X:
(3)
equals 1 for treated subjects and for untreated subjects. Hence, leaves the left-behind children sample unaltered, but weights the accompanied children sample such that it becomes representative of the left-behind children in terms of the distribution of X. In essence, IPWT balances X between treated and untreated and thus is a technique that mimics the characteristics of a randomized experiment. To test whether the data is indeed balanced after weighting the sample, the standardized difference of the mean of the propensity scores between left-behind children and accompanied children is calculated. In addition, a kernel density plot is estimated to assess the entire distribution of the propensity scores. Similar, standardized differences and kernel density plots are estimated for observed baseline characteristics, in order to test whether these are balanced between treated and untreated children (Morgan & Todd, 2008). Once the data is balanced, meaning that the propensity score model is correctly specified, the IPWT estimator will yield an unbiased estimate of the ATT:
∑ ∑
(4)
In equation (4) n is the total number of children in the sample.
To check the robustness of the estimates, three separate models are estimated for each dependent variable. The first model regresses the dependent variable on the treatment dummy, indicating parental migration status, including the constructed weights. The second model adds the dependent’s lagged variable from baseline (i.e. baseline
16 outcome variables), which is expected to be a good predictor of the outcome. The last model adds all baseline covariates from the propensity score model.
5.2. Assumptions
Rosenbaum and Rubin (1983) describe two assumptions underlying the propensity score analysis. The first is “no unmeasured confounders”, also referred to as the conditional independence assumption:
Assumption 1. Unconfoundedness:
This implies that conditional on a set of observed characteristics X, which are unaffected by the event of migration, potential outcomes are independent of treatment exposure. In other words, there are no unobservables that affect treatment assignment or the potential outcomes. Evidently, this is a strong assumption and cannot be formally tested. For instance, the validity of this assumption could be questioned when arguing that unobservables, such as parents’ enthusiasm, are likely to influence the decision to migrate and children’s cognitive ability scores. Trying to capture parents’ enthusiasm, I included pre-migration cognitive ability scores, and pre-migration mathematic and Chinese grades in the propensity score model. Throughout this thesis I will assume that this assumption holds.
A second requirement is the overlap or common support condition:
Assumption 2. Overlap:
The overlap condition says that the probability of being treated (i.e. being left-behind) or untreated (i.e. being accompanied) for every subject is strictly between zero and one, hence ruling out perfect predictability of treatment exposure given X. Assumptions 1 and 2 together are referred to as ‘strong ignorability’ and will yield an unbiased ATT estimate.
Furthermore, because I balance the data on baseline covariates including baseline outcome variables, a third assumption needs to be made – the parallel trend assumption:
17 Assumption 3. Parallel Trend: In absence of the intervention (i.e. parental migration) the treatment and control group would have parallel trends in the outcome variable.
In other words, the parallel trend assumption says that any remaining pre-migration differences between the treatment and the control group would have stayed constant over time if parental migration had not occurred. With two time periods, like is the case in this thesis, the parallel trend assumption is untestable. Therefore, throughout the rest of this thesis I assume this assumption to hold.
Moreover, I assume error terms to be serially uncorrelated, also referred to as the serial independence assumption. Finally, I make use of robust standard errors because of possible heteroskedasticity.
18 6. Results
6.1. Descriptive Statistics
The descriptive statistics of the dependent variables of the full sample are presented in Table 2. On average, children performed better on the number series test than on the word recall test. In addition, 86% of all children in the sample attended school in 2012. It should be noted that the majority of the children not in school was aged 15 or older, which is in line with China’s 9-year compulsory education law. Besides, the average current school level of all children was equal to junior high school. Furthermore, the average maximum education level in the household was between junior and senior high school.
Table 2.
Descriptive statistics of the outcome variables measured in 2012 (ages 12-18; N=1,215)
Dependent variable Mean SD N (%) M
Word recall test score (0-10) 6.00 1.80 1,152 (5.2) Number series test score (0-10) 6.32 2.34 914 (24.8) Log per capita family income (yuan) 8.24 1.32 1,178 (3.0) Log family expenditure on education for child i
(yuan)
7.50 1.37 1,026 (15.6)
School attendance 0.86 0.35 1,187 (2.3)
Current school level (1-7) 3.01 0.75 1,017 (16.3)
Maximum education level in HH (1-7) 3.41 0.86 1,215 (0.0)
Note. (%) M: percent missing data; current school level and maximum education level in HH range from 1
(illiterate/semi-illiterate) to 7 (master’s degree); the average current school level is junior high school; the average maximum education level in the household lies between junior and senior high school.
Source. China Family Panel Studies
6.2. Propensity Score Model
Children’s predicted propensity scores are estimated through a logit model, including all baseline characteristics described in part 4.2 Results of the logit model indicate the likeliness of having at least one migrant parent. Children in smaller families, who have at
2
The logit model excluded 3 observations, due to some missing values in one baseline covariate (mother’s years of schooling). Thus the remaining sample counts 1,212 rural children.
19 least one grandparent alive, who have less educated mothers and who live in poorer families, are more likely to have at least one migrant parent (see Table A2). It should be noted that there is no lack of overlap, since the maximum predicted propensity score in the treatment group is smaller than the maximum value in the control group. Therefore, all treated children have counterfactuals in the control group, and so the common support standards are satisfied without the need of trimming (Morgan & Todd, 2008). The predicted probabilities have a mean of 10.8, a standard deviation of 0.058 and ranges from 0.004 to 0.362. Moreover, the propensity score model correctly classifies 89.19% of the children into the treatment group, and thus does a relatively good job in predicting treatment exposure.
6.3. Weight and Balance Diagnostics
Prior to the balance diagnostics it is essential to describe the distribution of the created weights. The weights have a mean of 0.216, a standard deviation of 0.282, and ranges from 0.004 to 1. Austin (2015) argues that the IPTW method may be sensitive to large weights, however large weights are not of any concern in my study, as the maximum value equals 1. Furthermore, to assess whether the propensity score model is correctly specified, balance diagnostics are applied to the raw (i.e. before weighting) and the weighted sample.3 The absolute standardized difference of the means of the propensity scores between left-behind and accompanied children before weighting is 0.586, indicating significant imbalance as it exceeds the 0.10 threshold for balanced data (Austin, 2011b). The imbalance is further illustrated in Figure 1, where kernel density estimates show clear differences in the entire distribution of propensity scores of treated and untreated children in the unweighted sample. In addition, standardized differences of the baseline characteristics before weighting are reported in Table A3. It can be observed that 8 of the 17 baseline covariates have an absolute standardized difference larger than
3
It is difficult, if not impossible to verify whether the propensity score model is accurately specified. Therefore, the focus should be on assessing balance of observed covariates across treated and untreated children in the weighted sample (Austin & Stuart, 2015).
20 0.10, implying that these characteristics are imbalanced across treated and untreated subjects in the unweighted sample.
Figure 1. Kernel density estimates of treated and untreated children in the raw sample
Source. Author’s calculation
Left-behind children are younger, live in families with more generations, have less educated mothers, live in households where the maximum education level is lower, are poorer, live in communities where the agricultural population is larger, are more likely to attend school, and have lower math test scores compared to accompanied children. Therefore, it can be concluded that left-behind children and accompanied children are incomparable in the raw sample.
These steps are repeated after weighting the entire sample by the inverse probability of treatment. The weighted absolute standardized difference of the means of the propensity scores between treated and untreated equals 0.012, and is insignificant. As illustrated in Figure 2, the weighted distribution of propensity scores of the two groups is almost identical. Therefore, it can be concluded that after weighting the sample, the propensity scores are balanced across the treatment and the control group. The weighted standardized differences of the baseline covariates are also reported in Table A3. They can be interpreted as negligible, because none of the absolute differences exceed the threshold of 0.10 for balanced data.
21
Figure 2. Kernel density estimates of treated and untreated children in the weighted sample
Source. Author’s calculation
This implies that, after weighting, all baseline characteristics are balanced across the treated and untreated subjects. In addition, kernel density estimates of four continuous baseline outcome variables (i.e. word test score, math test score, log per capita family income, and log family expenditure on education for child i) and two discrete baseline outcome variables (i.e. maximum education level in the household and current school level) are illustrated in Figure 3 and Figure 4, respectively. It can be observed that the distributions of the baseline outcome variables are similar across treated and untreated children in the weighted sample. Consequently, systematic differences in measured baseline covariates between the left-behind and accompanied children have been reduced, indicating the effectiveness of the inverse probability of treatment weighting method.4 Subsequently, the effects of parental migration can be computed, by comparing outcomes of left-behind children with those of accompanied children in the weighted sample.
4
I strive to reduce systematic differences between treated and untreated subjects as much as possible, however it is impossible to entirely eliminate all differences between the two groups (Rubin, 2006).
22
Figure 3. Kernel density estimates of continuous baseline outcome variables of treated and untreated children in the weighted sample
23
Figure 4. Density estimates of discrete baseline outcome variables of treated and untreated children in the weighted sample
Note. Maximum education level and current school level are measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree) Source. Author’s calculation
24 6.4. Impact of Parental Migration
Estimates of the ATT are presented in Table 3 column 2. Left-behind and accompanied children do not significantly differ in their word recall test scores. In contrast, left-behind children score on average 0.58 points (p < 0.05) lower on the number series test than accompanied children. Yet, the negative effect is reduced to 0.40 points and loses its significance when explanatory baseline covariates are added to the model. Hence, there is evidence for an overall significant negative effect of parental migration on cognitive ability scores of left-behind children compared to non-migrant children.
Moreover, per capita family income significantly increased by 81% (p < 0.01) as a result of parental migration. The result is robust to adding explanatory baseline covariates, while the standard error declines. Surprisingly, migration of at least one parent negatively affected family expenditure on education relative to when both parents remain at home. Even though the result is insignificant, this finding is counterintuitive as a significant positive effect was expected and will be explained in the discussion. Alternatively, I tried to capture educational investment by children’s school attendance and current school level. No significant difference was found in the probability of school attendance of behind children relative to accompanied children. Neither do left-behind children and accompanied children significantly differ in their current school level, measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree). Finally as expected, parental migration significantly reduced the maximum education level in the household by 0.23 points (p < 0.01), measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree).
25 Table 3.
Average treatment effects on the treated in the full sample and by gender with accompanied children (of the same sex) as the control group Full sample (N=1,212) Girls (N=612) Boys (N=575)
N ATT Robust SE N ATT Robust SE N ATT Robust SE
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8) (9)
Word recall test score (0-10) 1,149 583 543
Without covariates 0.13 0.161 0.29 0.195 -0.05 0.260
With “word test score” from baseline 0.15 0.161 0.29 0.189 -0.04 0.261 With all baseline covariates from the PS model 0.13 0.154 0.29 0.183 -0.03 0.243
Number series test score (0-10) 911 460 432
Without covariates -0.58** 0.263 -0.37 0.345 -0.84* 0.433
With “math test score” from baseline -0.47* 0.249 -0.27 0.322 -0.75* 0.416 With all baseline covariates from the PS model -0.40 0.245 -0.38 0.311 -0.67* 0.378
Log per capita family income (yuan) 1,175 596 554
Without covariates 0.81*** 0.073 0.90*** 0.112 0.74*** 0.104
With dependent’s lagged variable from baseline 0.80*** 0.070 0.88*** 0.109 0.73*** 0.097 With all baseline covariates from the PS model 0.79*** 0.066 0.88*** 0.102 0.71*** 0.090 Log family expenditure on education for child i
(yuan)
1,023 513 502
Without covariates -0.18 0.133 -0.13 0.190 -0.26 0.194
With dependent’s lagged variable from baseline -0.14 0.125 -0.13 0.179 -0.22 0.182 With all baseline covariates from the PS model -0.14 0.110 -0.17 0.150 -0.24 0.157
26 Table 3. (Continued)
Note. * p < 0.1; ** p < 0.05; *** p < 0.01; current school level and maximum education level in HH are measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree)
Full sample (N=1,212) Girls (N=612) Boys (N=575)
N ATT Robust SE N ATT Robust SE N ATT Robust SE
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8) (9)
School attendance 1,184 597 562
Without covariates 0.08 0.287 0.08 0.409 0.04 0.413
With dependent’s lagged variable from baseline 0.09 0.301 0.12 0.459 0.04 0.413 With all baseline covariates from the PS model 0.15 0.296 0.29 0.475 0.08 0.403
Current school level (1-7) 1,014 514 491
Without covariates -0.08 0.072 0.02 0.107 -0.18* 0.094
With dependent’s lagged variable from baseline -0.07 0.053 0.00 0.077 -0.16** 0.073 With all baseline covariates from the PS model -0.05 0.039 -0.01 0.055 -0.13** 0.055
Maximum education level in HH (1-7) 1,212 612 575
Without covariates -0.23*** 0.071 -0.26** 0.108 -0.19* 0.097
With dependent’s lagged variable from baseline -0.24*** 0.049 -0.28*** 0.079 -0.19*** 0.063 With all baseline covariates from the PS model -0.23*** 0.045 -0.28*** 0.066 -0.19*** 0.061
27 7. Results by Gender
7.1. Preliminary Steps
To examine whether the effects of parental migration vary by children’s gender, the analysis is performed again, separately for girls and boys. So, left-behind girls are compared to accompanied girls and left-behind boys to accompanied boys. Hence, the full sample is split up into two samples. The first consists of 618 girls, of whom 10.7% is left-behind, and the second consists of 597 boys, of whom 10.9% is left-behind. The similar percentages indicate that the probability of having at least one migrant parent is not dependent on the gender of the child. The propensity score model is estimated for girls and boys separately, for which the exact same logit model is used as for the full sample (i.e. including the same baseline covariates). The results of the logit model, reported in Table A2, suggest that girls in smaller families, who have at least one grandparent alive, and who live in communities where the population in agriculture is larger, are more likely to be left-behind. In contrast, boys who have more educated fathers and who live in poorer families are more likely to have one migrant parent. While no trimming was needed for boys, three girls have been omitted from the sample, because their predicted propensity scores exceeded the highest value in the control group.5 All boys in the remaining sample attended school at baseline in 2010, while 94.9% of all the girls attended school in 2010. Besides, the propensity score models correctly classify 89.71% and 88.70% of the left-behind girls and boys in the treatment group, respectively. Furthermore, balance diagnostics are performed for girls and boys separately. Standardized differences and kernel density estimates of propensity scores and baseline covariates, before and after weighting, are presented in Figure A1, Figure A2, Figure A3, and Table A3. It can be concluded that weighting both samples by the inverse probability of treatment, results in balanced propensity scores and baseline covariates across treated and untreated subjects. Consequently, the outcomes of treated and untreated girls and boys can be compared in the weighted sample.
5In addition, the logit model for girls (boys) excluded 3 (22) observations, due to some missing
values in one (two) baseline covariate(s). Thus the remaining sample counts 612 (575) rural girls (boys).
28 7.2. Impact of Parental Migration by Gender
Column 5 and 8 in Table 3 report estimates of the ATT by gender. It can be observed that the effects of parental migration are different for girls and boys. Left-behind girls and accompanied girls do not significantly differ in their word recall test scores and their number series test scores. Therefore, the results seem to suggest that parental migration has an overall neutral effect on cognitive ability scores of left-behind girls relative to accompanied girls. Similar, left-behind boys and accompanied boys do not significantly differ in their word test scores. In contrast, left-behind boys score on average 0.84 points (p < 0.1) lower on the number test than accompanied boys. However, when explanatory baseline covariates are added to the model, the negative effect is reduced to 0.67 points (p < 0.1), but remains significant. Thus, it can be concluded that parental migration has an overall significant negative effect on cognitive ability scores of left-behind boys compared to non-migrant boys.
Furthermore, the following effects of parental migration are similar across left-behind girls and boys. Per capita family income significantly increased and family expenditure on education is negatively affected, but insignificant. Once more, the effect on education expenditure is counterintuitive; in fact a significant positive effect was expected, and will be explained in the discussion. Alternatively, I tried to capture educational investment by children’s school attendance and current school level. No significant difference was found in the probability of school attendance of left-behind girls and boys relative to accompanied girls and boys, respectively. Besides, left-behind girls are not affected in terms of their current school level. In contrast, parental migration decreased the current school level of left-behind boys by 0.18 points (p<0.1), measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree). Yet, the negative effect is reduced to 0.13 points (p<0.05) when explanatory baseline covariates are included, but gained significance, while the standard error declined. Finally as expected, parental migration significantly reduced the maximum education level in the household for both left-behind girls and boys.
29 8. Discussion
Drawing on two waves of data from the China Family Panel Studies and applying IPTW using predicted propensity scores; this study finds that parental migration decreases cognitive ability scores of left-behind children. This result is consistent with the finding obtained by Nguyen (2016) for left-behind children in India and Vietnam, in particular for long-term parental migration. The result can be explained by analyzing the economic and social mechanisms underlying the overall effect. On the one hand, migration of at least one parent significantly increased per capita family income, which is in line with the theoretical expectation, of higher earnings generated by migrant parents in urban regions. Although data on remittances is missing, the extra economic resources as an effect of parental migration are an indication that remittances are included the per capita family income measure. On the other hand, a positive effect of migration on family expenditure on education was expected, though a neutral effect is found. According to the CFPS 2012 questionnaire, family expenditure includes expenditures of family members who permanently live at home or who return home every week (CFPS, 2012). Thus, the expenditures of migrated parents are not included in this measure. A plausible explanation for the neutral effect on spending on education would be partial direct finance of education by migrant parents, rather than indirect finance of education through remittances to the left-behind family. Because of a lack of information on expenditures of migrants, I was unable to provide evidence for the latter in this study. Alternatively, I tried to capture educational investment by children’s school attendance and current school level. Though I found no significant results, Yang (2008) proves that increased economic resources, generated by migrant parents, increase educational expenditures in the Philippines. In summary, higher per capita family income illustrates the positive economic effect of parental migration. Nevertheless, the maximum education level in households with at least one migrated parent decreased significantly. This finding is as expected and illustrates the negative social effect of parental migration. Overall, left-behind children seem to be worse off than accompanied children in terms of their cognitive ability scores, which suggests that the negative social effect outweighs the positive economic effect of parental migration.
30 In addition, this thesis examined whether there is gender variance in the effect of parental migration. The results suggest that girls are not affected by the absence of at least one parent due to migration regarding their cognitive ability scores, whilst it has a detrimental effect on boys. For the interpretation of this result, it is worth referring to Table 1, which shows the parental migration status by the gender of the migrant parent. The majority of the left-behind children in my sample live in a father-only migrant household, meaning that in most cases children are separated from their father, while their mother remains at home. The absence of a father figure possibly affects left-behind girls and boys differently. While boys are separated from their father as role model, girls continue to be surrounded by their mother as role model. This interpretation of the results is in line with that of Wang (2014), who finds a negative effect of parental migration on school enrollment of boys, but a neutral effect on school enrollment of girls in rural China. Wang claims the absence of a father role to be the main cause of the negative effect on boys and suggests that mothers do not succeed in mitigating the adverse effect of father absence. Besides, a shift in head of household might affect girls and boys differently. The literature provides evidence that female head of households are beneficial for outcomes for girls but not for boys (Duflo, 2003). Hence, father absence and a shift in household decision-making authority are two possible explanations for the negative (neutral) link between parental migration and cognitive ability scores for boys (girls). However, Antman (2012) gives a different explanation to her results. Her finding that paternal domestic migration has no effect on educational attainment of children in Mexico, suggests that father absence does not play a crucial role in children’s educational outcomes. Yet, she finds that paternal migration from Mexico to the US increases girls’ educational attainment by one year, but does not significantly affect boys. The gender difference is explained by the stronger impact of remittances on educational investment in girls compared to boys. This argument can be extended to the results of my study. The strong male preference in rural China ensures financial constraint families to first invest in boys. Any extra financial resources are likely invested in the education of girls (Parish & Willis, 1993). Even though I was unable to provide evidence on increased educational investment in girls, the non-negative link between parental migration and cognitive
31 ability scores seems supportive of the hypothesis that girls benefit more from parental migration than boys.
There are two limitations to my research. First, I conditioned on children living in non-migrant families at baseline in 2010. As a result, 544 children who are left-behind by at least one parent by 2010 are not considered. Thus, I am unable to make a statement for these children. However, it would be interesting so see whether the results of this study are the same for the omitted group. Second, IPTW only allowed me to account for observed baseline variables. So, the ATT estimates may still be susceptible to bias induced by unmeasured confounders (Austin, 2011a; Austin, Mamdani, Stukel, Anderson, & Tu, 2005). However, in case test scores (i.e. Chinese grade, math grade, word test score, and math test score) at baseline capture unobservables (e.g. parent’s enthusiasm) that affect the decision to migrate and cognitive ability scores, it is plausible to argue that the inclusion of these pre-migration variables in the propensity score model reduce bias induced by unmeasured confounders. Nevertheless, I suggest future research to apply a combination of IPTW and differences-in-differences in order to control for time-invariant unobservables.
32 9. Conclusion
Drawing on data from two waves of the CFPS, this thesis has investigated the impact of parental migration on cognitive ability scores of left-behind children aged 12-18 in rural China. Parental migration is a worldwide phenomenon, but the largest rural to urban migration in human history has been witnessed in China. Millions of Chinese parents have left their children behind in the countryside due to restrictions imposed by the household registration system. The gross effect of migration on children is yet unknown, and may be twofold. For, on the one hand, children can benefit from the extra economic resources generated by migrant parents insofar as they are invested in them. Yet on the other hand, parental absence and a poorer domestic educational environment might disadvantage child development and children’s educational achievements. This thesis was the first study to apply IPTW using propensity scores to account for systematic differences in observed characteristics between left-behind and accompanied children. Conditioning on children living in non-migrant families at baseline, this study provides evidence on an immediate negative influence of parental migration on cognitive ability scores of left-behind children. The socio-economic mechanisms underlying the overall effect are illustrated by an increase in per capita family income and a decrease in the maximum education level in the household. The result suggests that the negative social effect outweighs the positive economic effect of parental migration.
Furthermore, there is evidence for gender variance in the effect of parental migration. In fact, the impact of parental absence due to migration is neutral for girls, but negative for boys. The gender difference can be explained in several ways. Firstly, the majority of the children in my sample are left behind by their father. Following the literature on child development, sons are more adversely affected by their father’s absence than girls are. Secondly, a shift in household decision-making authority from the father to the mother seems disadvantageous for sons, but not for daughters. Finally, considering the large male preference in rural China, the results can also be explained by a stronger impact of remittances on educational investments in girls. I suggest further research into the long-term impact of parental migration on left-behind children in China. This can be addressed once a nationwide comprehensive survey on left-behind children is available, which the Chinese government is planning to conduct in the near future.
33 References
Adams, R. H., & Page, J. (2005). Do international migration and remittances reduce poverty in developing countries?. World development, 33(10), 1645-1669.
All -China Women’s Federation, (2016). China to Conduct 1st Comprehensive Survey on
'Left-behind Children‘. Retrieved May 2, 2016, from
http://www.womenofchina.cn/womenofchina/html1/news/china/1603/2095-1.htm
Antman, F. M. (2012). Gender, educational attainment, and the impact of parental migration on children left behind. Journal of Population Economics,25(4), 1187-
1214.
Austin, P. C. (2011a). A tutorial and case study in propensity score analysis: an application to estimating the effect of in-hospital smoking cessation counseling on mortality. Multivariate behavioral research, 46(1), 119-151.
Austin, P. C. (2011b). An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate behavioral
research, 46(3), 399-424.
Austin, P. C., Grootendorst, P., & Anderson, G. M. (2007). A comparison of the ability of different propensity score models to balance measured variables between treated and untreated subjects: a Monte Carlo study. Statistics in medicine, 26(4), 734-
753.
Austin, P. C., Mamdani, M. M., Stukel, T. A., Anderson, G. M., & Tu, J. V. (2005). The use of the propensity score for estimating treatment effects: administrative versus clinical data. Statistics in medicine, 24(10), 1563-1578.
Austin, P. C., & Stuart, E. A. (2015). Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies. Statistics in medicine, 34(28),
3661-3679.
Brookhart, M. A., Schneeweiss, S., Rothman, K. J., Glynn, R. J., Avorn, J., & Stürmer, T. (2006). Variable selection for propensity score models. American journal of
epidemiology, 163(12), 1149-1156.
Calero, C., Bedi, A. S., & Sparrow, R. (2009). Remittances, liquidity constraints and human capital investments in Ecuador. World Development, 37(6), 1143-1154.
34 China Family Panel Studies. “2012 Questionnaire.” Survey. April 2012
Duflo, E. (2003). Grandmothers and granddaughters: old‐ age pensions and intrahousehold allocation in South Africa. The World Bank Economic
Review,17(1), 1-25.
Edwards, A. C., & Ureta, M. (2003). International migration, remittances, and schooling: evidence from El Salvador. Journal of development economics,72(2), 429-461. Falbo, T. (1991). The impact of grandparents on children's outcomes in China. Marriage
& Family Review, 16(3-4), 369-376.
Institute of Social Science Survey, Peking University, 2015, "China Family Panel Studies (CFPS)", http://dx.doi.org/10.18170/DVN/45LCSO, Peking University Open Research Data Platform, V3
Liang, Z., Guo, L., & Duan, C. (2008). Migration and the well-being of children in China. The Yale-China Health Journal, 5, 25-46.
Lu, Y. (2012). Education of children left behind in rural China. Journal of Marriage and Family, 74(2), 328-341.
Lu, Y. (2014). Parental Migration and Education of Left‐ Behind Children: A Comparison of Two Settings. Journal of Marriage and Family, 76(5), 1082-1098.
McKenzie, D., & Rapoport, H. (2011). Can migration reduce educational attainment? Evidence from Mexico. Journal of Population Economics, 24(4), 1331-1358. McLanahan, S., & Sandefur, G. (1994). Growing Up with a Single Parent. What Hurts,
What Helps. Harvard University Press, 79 Garden Street, Cambridge, MA 02138.
Morgan, S. L., & Todd, J. J. (2008). A diagnostic routine for the detection of consequential heterogeneity of causal effects. Sociological Methodology,38(1),
231-281.
National Bureau of Statistics. (2016). China’s Economy Realized a Moderate but Stable and Sound Growth in 2015. Retrieved May 2, 2016, from http://www.stats.gov.cn/english/PressRelease/201601/t20160119_1306072.html
Nguyen, C. V. (2016). Does parental migration really benefit left-behind children? Comparative evidence from Ethiopia, India, Peru and Vietnam.Social Science &
35
Parish, W. L., & Willis, R. J. (1993). Daughters, education, and family budgets Taiwan experiences. Journal of Human Resources, 863-898.
Parreñas, R. S. (2005). Children of global migration: Transnational families and
gendered woes. Stanford University Press.
Rosenbaum, P. R. (1987). Model-based direct adjustment. Journal of the American
Statistical Association, 82(398), 387-394.
Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41-55.
Rubin, D. B. (2006). Matched sampling for causal effects. Cambridge University Press. Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology, 66(5), 688.
Santrock, J. W. (1972). Relation of type and onset of father absence to cognitive development. Child Development, 455-469.
Thomas, D. (1994). Like father, like son; like mother, like daughter: Parental resources and child height. Journal of human resources, 950-988.
Wang, S. X. (2014). The effect of parental migration on the educational attainment of their left-behind children in rural China. The BE Journal of Economic Analysis &
Policy, 14(3), 1037-1080.
World Bank. (2016). Graph urban population. Retrieved from
http://data.worldbank.org/indicator/SP.URB.TOTL.IN.ZS?end=2015&locations= CN&start=1980
Xie, Y., & Hu, J. (2014). An introduction to the China family panel studies (CFPS). Chinese Sociological Review, 47(1), 3-29.
Xu, H., & Xie, Y. (2015). The causal effects of rural-to-urban migration on children’s well-being in China. European sociological review, jcv009.
Yang, D. (2008). International migration, remittances and household investment: Evidence from Philippine migrants’ exchange rate shocks. The Economic
Journal, 118(528), 591-630.
Zhang, H., Behrman, J. R., Fan, C. S., Wei, X., & Zhang, J. (2014). Does parental absence reduce cognitive achievements? Evidence from rural China. Journal of Development Economics, 111, 181-195.
36 Appendix
Table A1.
Descriptive statistics of the baseline covariates (pre-migration measures from 2010) (N=1,215)
Covariate Mean SD Min Max
Age 14.83 1.97 12 18
Male 0.49 0.50 0 1
Family size 4.93 1.42 3 12
Generations 2.41 0.52 2 4
Father’s years of education 5.53 4.05 0 15
Mother’s years of education 3.65 3.87 0 15
Maximum education level in HH (1-7) 3.27 0.89 1 7 Log per capita family income (yuan) 7.88 1.75 0 10.99
Population in agriculture (%) 49.84 24.56 0 100
Travel time to provincial capital (hours) 5.49 5.07 0 48
Attend school 0.96 0.20 0 1
Current school level (1-7) 2.30 0.73 0 4
Log family expenditure on education for child i (yuan) 5.44 2.22 0 9.98
Chinese grade in last exam 69.11 30.90 0 135
Math grade in last exam 65.73 33.72 0 146
Word test score 20.78 8.07 0 34
Math test score 10.70 4.82 0 24
Note. Maximum education level and current school level range from 1 (illiterate/semi-illiterate) to 7
(master’s degree); the average maximum education level in the household lies between junior and senior high school; the average current school level lies between primary school and junior high school.
37 Table A2.
Results of logit model (parental migration status regressed on all baseline covariates)
Full sample (N=1,212) Girls (N=615) Boys (N=575) SE SE SE Age -0.021 0.079 -0.093 0.115 0.072 0.118 Male -0.040 0.196 - - - - Family size -0.222** 0.097 -0.326** 0.150 -0.121 0.137 Generations 0.485** 0.229 0.611* 0.341 0.405 0.326
Father’s years of schooling 0.038 0.029 0.010 0.039 0.088** 0.044 Mother’s years of schooling -0.060** 0.030 -0.060 0.042 -0.068 0.045
Maximum education in HH
(1-7) -0.132 0.147 -0.079 0.203 -0.261 0.222
Log per capita family income
(yuan) -0.392*** 0.125 -0.299 0.182 -0.442** 0.182 Population in agriculture (%) 0.007 0.004 0.013** 0.006 0.001 0.007
Travel time to provincial capital
(hours) -0.001 0.021 -0.015 0.038 0.008 0.026
Attend school 2.131 1.502 2.028 1.917 - -
Current school level (1-7) -0.123 0.312 -0.044 0.438 -0.378 0.466 Log family expenditure on
education for child i (yuan) -0.045 0.081 -0.085 0.122 -0.007 0.113 Chinese grade in last exam -0.002 0.008 0.008 0.014 -0.006 0.011 Math grade in last exam -0.005 0.006 -0.009 0.009 -0.004 0.009 Word test score 0.014 0.018 0.007 0.027 0.011 0.026 Math test score -0.011 0.035 0.039 0.052 -0.052 0.049
Note. * p < 0.1; ** p < 0.05; *** p < 0.01; maximum education level and current school level are measured on a scale of 1 (illiterate/semi-illiterate) to 7 (master’s degree)
38 Table A3.
Balance of baseline covariates of the treated and the control in the full sample and by gender in the raw and weighted sample
Full sample (N=1,212) Girls (N=612) Boys (N=575)
Covariate Mean T Mean C St. Dif. Mean T Mean C St. Dif. Mean T Mean C St. Dif.
Age R 14.53 14.87 -0.179* 14.78 14.90 -0.065 14.35 14.85 -0.258* W 14.53 14.53 -0.001 14.78 14.71 0.037 14.35 14.37 -0.009 Male R 0.50 0.49 0.011 0.00 0.00 - 1.00 1.00 - W 0.50 0.49 0.004 0.00 0.00 - 1.00 1.00 - Family Size R 4.88 4.93 -0.042 4.90 5.11 -0.150* 4.88 4.74 0.102* W 4.88 4.88 -0.001 4.90 4.90 0.004 4.88 4.89 -0.011 Generations R 2.47 2.40 0.134* 2.44 2.40 0.090 2.49 2.41 0.165* W 2.47 2.48 -0.014 2.44 2.46 -0.035 2.49 2.50 -0.022 Father’s years of education R 5.31 5.56 -0.060 5.10 5.52 -0.106* 5.49 5.57 -0.019
W 5.31 5.30 0.003 5.10 5.13 -0.009 5.49 5.43 0.016 Mother’s years of education R 2.77 3.76 -0.261* 2.95 3.74 -0.198* 2.63 3.80 -0.322*
W 2.77 2.79 -0.006 2.95 2.92 0.008 2.63 2.66 -0.007 Maximum education level in HH R 3.10 3.29 -0.220* 3.17 3.31 -0.157* 3.03 3.27 -0.272*
W 3.10 3.09 0.005 3.17 3.15 0.025 3.03 3.02 0.007 Log per capita family income (yuan) R 7.56 7.92 -0.196* 7.71 7.84 -0.078 7.47 8.00 -0.272*
W 7.56 7.57 -0.003 7.71 7.64 0.044 7.47 7.50 -0.017 Population in agriculture (%) R 54.11 49.33 0.199* 55.60 47.82 0.326* 53.54 50.88 0.110*
