Evolution of family planning policy in China

Chinese society stabilised and experienced sizable population growth after 1949, and the TFR reached more than six births per woman in the early 1960s (Reference BanisterBanister, 1987). It directly pushed China’s central government to issue an administrative document regarding population control in 1962, which marked the beginning of China’s family planning policies.

During the 1970’s, in accordance with this document, Chinese local authorities suggested the number of children that each couple could have, be dropped from three to two. Simultaneously, a campaign arose for later marriage, longer intervals between births, and fewer children, known as the ‘Later, Longer, and Fewer’ policy. This mostly voluntary campaign was successful and led to a massive decrease in the TFR.

However, overpopulation fears persisted as the baby boomers born after the great famine were entering reproductive ages. Then, recognised as the world’s strictest family planning policy, the one-child policy was implemented in 1979. It requested that each Han Chinese couple only have one child, and households with ‘above-quota births’ would be penalised. A second child was permitted only under specific circumstances.

Owing to the traditional preference for boys, the rule was deemed virtually unenforceable. The one-child policy was gradually relaxed to allow couples in rural areas, where around 70% of the total population lived, to have a second birth if the first one was a girl—called the one-and-a-half-child policy. After 1982, more provinces relaxed the conditions for a second birth, such as if both parents were the only children or if the first child was disabled.

China’s fertility policy has become relatively stable, and since 1990, the TFR has dropped lower than the replacement level. In recent decades, China has begun to face the mounting pressures associated with continued low fertility, and the one-child policy in such a demographic contest was no longer defensible (Reference PengPeng, 2011). The Chinese government has begun to relax the population control policy gradually. Around 2014, couples in which one partner was an only child were generally allowed to have a second child, and the universal two-child policy replaced the iconic one-child policy on January 1, 2016.

Due to the universal two-child policy, the sluggish fertility trend has not been reversed as expected. The number of newborns in China was 17.86 million in 2016, reaching the maximum level since 2000 and decreasing in later years. Considering the fading fertility accumulation effect and women’s lukewarm willingness to give birth a second time, the number of newborns has plateaued. Having only one child is endogenously determined within households, not by the policy constraints. Knowing the possible barriers to child-bearing and taking feasible action is the top priority for the Chinese government.

Literature review

Previous literature has revealed that having children leads to a reduction in the female labour supply (Reference Angrist and EvansAngrist and Evans, 1996). This negative effect not only arises among women aged 20-39 but also persists among older women (Reference Bloom, Canning, Fink and FinlayBloom et al., 2009). The impact of children on women’s labour market outcomes is widespread, affecting in terms of labour force participation, working hours, wages, occupations, sectors, and firms, while men are virtually unaffected (Reference LokshinLokshin, 2004; Reference Kleven, Camille and SøgaardKleven et al., 2019). Furthermore, mothers have a lower transition rate to re-employment than childless women (Reference Frodermann and MüllerFrodermann et al., 2019).

Several fruitful studies have examined the aforementioned ‘motherhood penalty’. Women who take a break from work for childbearing experience downward occupational mobility, long-term wage decline, and a low probability of being re-employed, even for only a few months (Reference Angrist and EvansAngrist and Evans, 1996). Reference WaldfogelWaldfogel (1997) finds that, compared to unmarried women, hourly wages are 6% lower for women with one child and 13% for women with two children. The annual income gap is even larger because mothers are more inclined to seek part-time jobs or switch their work departments in the process of parenting, and such translations often lead to lower wages (Reference Adair, Guilkey and BisgroveAdair et al., 2002; Reference Juhn and McCueJuhn and McCue, 2017). Reference Yu and XieYu and Xie (2014) find that in China having one child reduces women’s wages by 7.0%, and two children reduce them by 16.8%, compared to childless women. Reference Budig and EnglandBudig and England (2001) summarise the four causes of mothers’ declining income: childbirth and childcare lead to (1) interruptions in work experience or at least interrupt their full-time work experience; (2) less efficient work; (3) choice of ‘mother-friendly’ jobs rather than highly-paid jobs; (4) experience discrimination from employers.

Role Conflict Theory asserts that this situation is due to the inherent tension between household labour and career development. Women undertake the roles of mothers and employees, which induces virtually exclusive family-work conflict (Reference Greenhaus and BeutellGreenhaus and Beutell, 1985). Throughout the world, the general social culture, regardless of Western or Eastern, agrees that women’s responsibility for taking care of the family is a greater priority than their career development (Reference FursmanFursman, 2002). In East Asian countries, such as China, Japan, and South Korea, mothers ‘opt out’ of the labour market because of gender discrimination, limited childcare options, gender-specific role expectations for women, and male-centered employment systems (Reference Kinoshita and GuoKinoshita and Guo, 2015; Reference Rudolf and KangRudolf and Kang, 2015). Reference He and ZhuHe and Zhu (2015) exploit twin births as a natural experiment and find a negative, but minimal, causal impact of a second child on the female labour force participation in urban China. They infer that relaxing the one-child policy does not significantly affect the urban female labour supply. Reference Kinoshita and GuoKinoshita and Guo (2015) point out that since childcare is generally provided by grandparents in rural areas, reduction in fertility does not increase parental labour supply. Due to the extended family structure, the researchers suggest a small or even non-existent relationship between fertility and the female labour supply in China, which may otherwise exist in developed countries. They infer that the effect of the two-child policy on the labour supply, if any, tends to be minimal. Reference Agarwal, Li, Qin and WuAgarwal et al. (2019) study the effect of selective two-child policy (if one parent is from an only-child family, the couple is allowed to have a second child) on the female workforce and find a significantly negative influence.

DID method

This paper employs the classic difference-in-differences (DID) method (Reference Angrist, Krueger, Ashenfelder and CardAngrist and Krueger, 1999), which has been widely utilised in public policy evaluations in recent years, to avoid potential endogenous problems, such as omitted variables or causality, and to improve estimation accuracy.

Considering that the universal two-child policy has been legally effective since 2016, we use the sample of 2014 China Family Panel Studies (CFPS) data as the pre-treatment group, with the 2018 data as the post-treatment group. The greatest challenge in our study is that the treatment group is not distinct. That is, there is no simple distinction between treatment and no treatment, as would be the case in most situations in which a DID method might be applied. This challenge calls for further exploration. Owing to endogenously determined local regulations and variations in one-child policy implementation across provinces (Reference McElroy and YangMcElroy and Yang, 2000), the universal two-child policy’s influence differs by region. If the province had a lower actual fertility rate before the universal two-child policy came into effect, it would probably be more strongly affected by this new policy. If the province’s fertility rate was already higher, it could be less influenced, so we need to find a feasible fertility rate as the indicator to divide the sample into the treatment and control groups. Finally, we select the policy fertility rate (PFR). We set aside the commonly used TFR in regions because scholars have debated its accuracy in China (e.g., Reference Chen and YangChen and Yang, 2014; Reference GuoGuo, 2011; Reference ZhuZhu, 2012). The PFR is defined as the average number of children born if a woman gives birth according to the local policy during her lifetime (Reference Guo, Zhang and GuGuo et al., 2003). Local authorities’ implementation variables, such as local-specific situations, fines for above-quota births, one-child subsidies, and the provision of contraceptives, have often varied across regions; so the PFR differs in each province.

The population-weighted national PFR of 1.465 is used as the cut-off rate based on the calculations in Table 1, following the identification strategy by Reference DufloDuflo (2001). After eliminating Tibet^{Footnote 1}, we group 12 provinces as the treatment group (Jiangxi, Shandong, Jilin, Guangdong, Heilongjiang, Liaoning, Chongqing, Sichuan, Tianjin, Beijing, Jiangsu, Shanghai) and 13 provinces as the control group (Hubei, Zhejiang, Hunan, Anhui, Fujian, Shanxi, Henan, Shaanxi, Guangxi, Gansu, Hebei, Inner Mongolia, Guizhou, Yunnan, Qinghai, Ningxia, Hainan, Xinjiang).^{Footnote 2}

Table 1. PFR of China’s 30 provinces (population-weighted corrected).

Source: Reference Guo, Zhang and GuGuo, Zhang, et al. (2003).

According to the Statistical Yearbooks in each province, growths in the birth rate for the treatment group are much higher than those for the control group after the universal two-child policy. Taking 2014 as a reference, growths in the birth rate are 7.38%, 10.45%, and 0.35% for the treatment group, and 1.37%, 5.16%, and −6.44% for the control group in 2016, 2017, and 2018, respectively, as shown in Figure 1. The growths in birth rate differ significantly between the control and treatment groups, and the latter is much more sensitive to the universal two-child policy. The assumption is affirmed to be justified and valid.

Figure 1. Annual rate of growth of birth rate in China, 2015–2018.

Source: Calculated by the data from China Statistical Yearbook of 2014–2018.

In practice, we use the following regression-adjusted DID model to examine the policy impact. Consider

(1) $Em{p}_{it}={\alpha }_0+{\alpha }_1{T}_i+{\alpha }_{2}Trea{t}_i+{\alpha }_{3}Trea{t}_i\times T_i+\alpha X_{it}+{\theta }_j+{\epsilon }_{it},$

(2) $ln{Y}_{it}={\alpha }_0+{\alpha }_1{T}_i+{\alpha }_{2}Trea{t}_i+{\alpha }_{3}Trea{t}_i\times T_i+\alpha X_{it}+{\theta }_j+{\epsilon }_{it},$

where $i,t,j$ denote the individual, year and province, respectively; $Em{p}_{it}$ denotes the employment dummy, which equals 1 if a woman $i$ in year $t$ is employed, while it equals 0 if she doesn’t have a job. ln $Y_{it}$ denotes the logarithm of total labour incomes per month in renminbi (CNY) from her primary job. Price adjustments are made to nominal wages in 2014 and 2018, using the 2010 price index as a benchmark. Also, the original data is two-sided 1% winsorised to eliminate outliers. $Trea{t}_i$ is the treatment group dummy, which equals 1 if the individual comes from the treatment group, with a lower PFR before the universal two-child policy. $T_i$ is the policy timing dummy, which equals 1 if the person is in the post-treatment group, that is, the 2018 CFPS database used in this article. These two variables, $Trea{t}_i$ and $T_i$ , pick up the two main effects of being in a treatment group and after the two-child policy. Essentially, the coefficient of the interaction term is our DID estimator, which measures the employment or income gap attributed to the affirmative universal two-child policy. Because the omitted variables affect the effectiveness of the DID method (Reference MeyerMeyer, 1995), we add a vector of variables $X_{it}$ in the equation to control for some demographic characteristics that also contribute to the employment and income, such as age, age square, education level, marital status, health, and enterprise ownership; ${\theta }_j$ is the dummy variable of province $j$ and ${\epsilon }_{it}$ is the error term. The robust standard errors are clustered at the individual level.

Heckman two-step estimation

Applying linear models such as ordinary least squares (OLS) leads to a bias in the income estimation because the observations’ distribution is conducted selectively or not randomly chosen (Reference HeckmanHeckman, 1976). Specifically, we can only follow women’s income when they are hired, so the variable ln $Y_{it}$ does not satisfy the DID method’s randomness requirement. A Heckman two-step estimation is strongly recommended to rectify the non-random selection bias and mitigate the effects of heteroscedasticity.

In the selection stage, a Probit model would be conducted. Consider

(3) $P\left(Em{p}_{it}=1\right)={\alpha }_0+{\alpha }_1{T}_i+{\alpha }_{2}Trea{t}_i+{\alpha }_{3}Trea{t}_i\times T_i+\alpha X_{it}+{\theta }_j+{\epsilon }_{it}$

where P(Emp_it = 1) indicates the probability for a woman to be employed, and other variables are the same as in equation (1).

At the outcome stage, an OLS regression applies to estimate the relationship and the extent of the independent variables constrained by the Probit model results in the first stage. We calculate the Inverse Mills Ratio (IMR) and put it into equation (2) as a correction term. This gives us

(4) $ln{Y}_{it}={\beta }_0+{\beta }_1{T}_i+{\beta }_{2}Trea{t}_i+{\beta }_{3}Trea{t}_i\times T_i+\beta Z_{it}+\lambda IM{R}_i+{\theta }_j+{\epsilon }_{it}$

where a vector of variables $Z_{it}$ controls the demographic characteristics that affect women’s incomes. Considering that work experience and different industries largely affect earnings, new variables, like work experience, work experience square, and industry, are introduced. Furthermore, to meet the exogeneity requirement of the Heckman two-step estimation, we take the variable ‘marital status’ as the exclusive variable, which is only in the Probit model of the first stage and is omitted from $Z_{it}$ in the outcome stage. The reasons are as follows.

First, it is an exogenous variable, and it doesn’t impair the IMR. Second, it is a conventional independent variable for explaining female labour participation (e.g., Reference Eckstein and LifshitzEckstein and Lifshitz, 2011; Reference Juhn and McCueJuhn and McCue, 2017; Reference Killingsworth, Heckman, Ashenfelter and LayardKillingsworth and Heckman, 1986). Third, ‘marital status’ does not directly affect people’s income, and has not been considered in many income studies such as the famous Mincer wage equation (Reference MincerMincer, 1974). Therefore, it is reasonable to omit ‘marital status’ in the second stage and use the IMR to reflect its possible indirect influence.

Data description

The China Family Panel Studies (CFPS) program was launched in 2000 and is run every 2 years. The surveys exist at three levels: community, family, and individual levels, covering 25 provinces and more than 90% of the population in China.

Prior to the quantitative study, we restrict our participants to women between 16 and 54 years old who are neither full-time students nor incapacitated for work. We set 16 as the lower bound because it is the statutory age for work in China and employed women under 16 are rare. We use 54 as the upper bound according to the current retirement age for women.^{Footnote 3} In addition, since agricultural production is quite different from formal employment in urban areas, those who have rural hukou (household registration) are eliminated from the sample. We obtain 3010 and 2095 individuals in 2014 and 2018. Specifically, we consider the following variables:

Due to space limitations, the summary statistics of control variables are omitted from this paper and available on request.

Baseline DID method

In column 1 of Table 2, the coefficient of interaction ${\mathrm{Treat}}_i\times T_i$ is negative and significant at the 10% statistical level, demonstrating that the universal two-child policy has significantly decreased female labour participation. The switch from the one-child policy to the universal two-child policy reduces female employment by as much as 4.06%.

Table 2. Effect of the universal two-child policy on female employment and income.

Note: For the variable education level, the reference group is ‘junior high school or below’; for enterprise ownership, the reference is ‘other’ ownership; for industry, the reference group is ‘Public administration and social organisation’; and for health status, the reference is the ‘very healthy’ group. Due to space limitations, we only report the regression results for employees who have a bachelor’s degree or above, are unhealthy, work in foreign/Hongkong/Macao/Taiwan enterprises from the manufacturing industry. The complete results are available by request. Clustered standard errors are in parentheses. *, **, and *** represent significance levels of 10%, 5% and 1%, respectively.

Childcare reduces the female labour supply. Under the universal two-child policy, many households decide to have a second child, which can compete with working hours for young women. Furthermore, the statutory admission for kindergarten is children over 3 years old, and few government-funded childcare or affordable childcare services are available for children under 3. Mothers have no choice but to exit the labour market if they have no way to share the burden of childcare with their parents or the parents of their spouses.

The control variables can also be attributed to our model’s validity and accuracy, which have substantial economic implications. First, the positive coefficient of age and the negative coefficient of age square indicate that: as women get older, the probability that they stay in the labour market will rise but at a decreasing rate, and once age exceeds a certain threshold, the employment probability declines. Second, marriage has a significantly negative effect on female employment. Third, the coefficient of education level is significantly positive, denoting that more education is beneficial to female employment. A bachelor’s degree or above boosts women’s employment rate by 11.03%, compared to those who graduate from junior high school or below. Furthermore, among these six different ownership types, foreign enterprises (including Hong Kong, Macao and Taiwan) improve women’s employment by 5.35% at the statistical level of 10%, compared to ‘others’.

Heckman two-step estimation

Since sample selection bias cannot be ignored in our study, the Heckman two-step estimation is exploited, and the result is presented in column 2 of Table 2.

After controlling for individual characteristic variables, the coefficient of IMR is significant at the 1% statistical level, confirming that selective bias does exist and the result from the simple OLS is unreliable. The coefficient of $Trea{t}_i\times T_i$ means that the universal two-child policy has decreased women’s monthly income by 10.43% ( $e^{-0.1101}-1$ ) at the 10% level. After eliminating bias, we develop the confidence to attribute this decline to the universal two-child policy.

Observing other control variables, the coefficients of female age and age squared indicate that an inverted U-shaped relationship exists between age and income, which is consistent with the classical Mincer wage equation (Reference MincerMincer, 1974). Work experience and experience square denote an inverted U-shaped relationship as well. Thus, receiving more education could considerably improve women’s earnings. Women working in government agencies, public institutions, state-owned enterprises, and foreign enterprises (including Hong Kong, Macao and Taiwan) earn significantly more money than women working for ‘other employers’. Regarding health status, self-evaluation of unhealthy status significantly decreases female income by 28.29% ( $e^{-0.3327}-1$ ) at the 1% level compared to a very healthy status.

Parallel-trend assumption

Utilisation of the DID estimation must satisfy the parallel-trend assumption, which requires the control and treatment group to demonstrate parallel and consistent trends before the exogenous shock occurs.

We use the event study^{Footnote 4} method to test the policy’s exogeneity. We take the year 2012 as a reference group and use $Yea{r}_i^{\tau }$ is a dummy variable to represent the year before and after the policy. ‘ $Trea{t}_i\times T_i$ ’ is replaced by ‘ $Trea{t}_i\times Yea{r}_i^{\tau }$ ’ in equations (1) and (4), and the expanded DID model is shown below:

(5) $Em{p}_{it}={\alpha }_0+T_i+Trea{t}_i+\sum _{\tau }{\alpha }_{\tau }Trea{t}_i\times Yea{r}_i^{\tau }+\alpha X_{it}+{\theta }_j+{\epsilon }_{it}$

(6) $ln{Y}_{it}={\beta }_0+T_i+Trea{t}_i+\sum _{\tau }{\beta }_{\tau }Trea{t}_i\times Yea{r}_i^{\tau }+\beta Z_{it}+{\theta }_j++\lambda IM{R}_i+{\epsilon }_{it}$

where $Yea{r}_i^{\tau }$ ( $\tau$ =14, 2016, 2018) is the dummy variable and equals 1 if the year is $\tau$ and 0, otherwise.

Columns 1 and 2 of Table 3 show that the coefficient of ${\mathrm{Treat}}_i\times Yea{r}_i^{2014}$ is not significant, while ${\mathrm{Treat}}_i\times Yea{r}_i^{2016}$ and ${\mathrm{Treat}}_i\times Yea{r}_i^{2018}$ are significant. That is, the two sets of samples are comparable, and the new fertility policy is exogenous.

Table 3. Paralled-trend test and dynamic effect.

Note: Due to space limitations, only interaction terms’ coefficients are reported, and the control variables are available by request. Clustered standard errors are in parentheses. *, **, and *** represent significance levels of 10%, 5% and 1%.

Figures 2 and 3 show the results of the parallel-trend test more apparently. They reveal that, before the implementation of the universal two-child policy in 2016, the interaction coefficient is not significantly different from 0 (the 95% confidence interval contains a value of 0) but is significantly negative after the policy. This finding indicates that both groups are indeed parallel and consistent before the policy in the female employment or income estimation.

Figure 2. Parallel-trend test for female employment.

Figure 3. Parallel-trend test for female income.

Dynamics of the policy effect

We add the data from 2016 to examine the dynamics of this policy effect. Specifically, we replace $t_i$ in equations (1) and (4) with dummy variables $Yea{r}_i^{2016}andYea{r}_i^{2018}$ Columns 3 and 4 of Table 3 report the estimations. The universal two-child policy decreases female employment by 3.93% and 7.15% in 2016 and 2018, respectively. We find that the coefficient increases during the two years, indicating that the policy’s effect on labour participation remains strong over time.

The policy effect on income is not significant in 2016, whereas it significantly declines by 10.61% ( $e^{-0.1122}-1$ ) in 2018, as shown in column 4. Hence, there is a time lag between policy implementation and income change. For instance, it might take quite a while to decrease or increase employees’ salaries due to the rigidity of the salary payments system in a company.

Resample test

A potential problem in our research is whether the designated year impairs our conclusion. Since the universal two-child policy came into force in 2016 and the CFPS database is updated to 2018, we cannot find another post-treatment cohort except, 2018. Therefore, we reselect the individuals in 2012 as the pre-treatment group and conduct the regressions again. Columns 1 and 2 of Table 4 show that the interaction coefficients are still significantly negative, and the absolute values are much higher than the benchmark outcomes. That is probably because, before the universal two-child policy, most provinces in China began to implement the two-child fertility policy for couples in cases in which either the husband or the wife is from a single-child family in 2014. After we choose individuals in 2012, to a large extent, we have picked up the effects of these two fertility policies on female labour market outcomes, which are both negative.

Table 4. Robustness test.

Note: Due to space limitations, only the interaction terms’ coefficients are reported, and the control variables are omitted but available by request. Clustered standard errors are in parentheses. *, **, and ***represent significance levels of 10%, 5% and 1%.

Placebo test

The division of the treatment and control groups plays a crucial role in our study. A placebo test is conducted by constructing a false treatment group and a false control group. In the previous study, the control and treatment groups are divided based on the PFR of 1.465. That is, we include the province with a PFR below 1.465 in the treatment group, and others in the control group. To construct the false groups, we number these 25 provinces and sort them by PFR from smallest to largest. We pick out the odd-numbered provinces as the false treatment group and the even-numbered provinces as the false control group.^{Footnote 5} Therefore, before the universal two-child policy is set, the average PFRs of the two groups are not evidently different from each other.

The results in columns 3 and 4 of Table 4 show that, the coefficients of the false treatment group and the post-treatment period, that is, ${\mathrm{Treat}}_i\times T_i$ , are insignificant. This finding indicates that, if we change the treatment group, we cannot obtain similar results consistent with Table 2, thus proving that the conclusions of Table 2 are robust and reliable.

PSM-DID test

This subsection selects individual characteristic variables for propensity score matching (PSM) to control the bias induced by choosing the two groups. We employed Logit estimation and 1:4 nearest neighbour matching. The vector of variables $X_{it}$ including age, age squared, marital status, education level, enterprise ownership, and health is used in the employment estimation. ‘Industry’, ‘work experience’, and ‘work experience square’ are added to the vector to assess female income. There is no significant difference after PSM, and the bias of control variables is eliminated, since the p-values of the t-tests are greater than 10%.^{Footnote 6}

The PSM-DID estimations are shown in Columns 5 and 6 of Table 4, which are consistent with our previous conclusion: the universal two-child policy negatively affects the employment and income of women who live in cities and towns.

Introduction of the macro variable

The control variables in the regressions are microdata at the individual level, but labour market outcomes are also substantially influenced by the macroeconomic environment. We added the GDP growth rate in each province to reduce macroeconomic interference. The results do not evidently change the size and significance of the coefficients in columns 7 and 8 of Table 4, and we believe the benchmark regression model is reliable from this perspective.

Based on age

We divided the urban female workforce into four groups based on age: under 25 years old, 25 to 29 years old, 30 to 34 years old, and 35 to 39 years old.^{Footnote 7} Then we performed the estimations using the DID method and Heckman two-step estimation within each group.

Column 1 in Table 5 shows that the interaction coefficient is surprisingly low at −23.99%; that is, the universal two-child policy has brought a 23.99% decrease in women’s labour participation under 25. The policy has no significant effect on female employment or income in other age groups. Thus, we have the confidence to assert that the universal two-child policy has a rather negative impact on younger females.

Table 5. Heterogeneity analysis.

Note: Due to space limitations, only interaction terms’ coefficients are reported, and the control variables are omitted but available by request. Clustered standard errors are in parentheses. *, **, and *** represent significance levels of 10%, 5% and 1%.

Based on education

Education level holds great importance in people’s employment and income generation, and we divided education into four groups.

The results in columns 3 to 6 of Table 5 indicate that the universal two-child policy has decreased higher-educated women’s income by 29.59% ( $e^{-0.3509}-1$ ), which is significant at the 1% statistical level. Although receiving more education can improve female earnings as discussed in Table 2, the universal two-child policy’s implementation leads to a more considerable income decline for women with higher academic qualifications. The ‘motherhood penalty’ among better educated or high-skilled women is more severe than among the lower-educated (Reference Anderson, Binder and KrauseAnderson et al., 2002; Reference Yu and XieYu and Xie, 2014).

Based on enterprise ownership

After carrying out the regression within each subgroup, it was found that only ‘private companies/individual businesses’ have a significant adverse relationship with female employment.^{Footnote 8} The interaction coefficient is −6.93% and is significant at the 10% level. After the new policy, women find it more challenging to be hired at a private enterprise, and this is probably because the private companies are more cost-sensitive.