Journal of Comparative Economics 50 (2022) 713–731

Contents lists available at ScienceDirect

Journal of Comparative Economics

journal homepage: www.elsevier.com/locate/jce

The timing of first marriage and subsequent life outcomes: Evidence

from a natural experiment✩
Yi Chen a ,∗, Yi Zhao b
a ShanghaiTech University, School of Entrepreneurship and Management, China
b Tsinghua University, School of Economics and Management, China

ARTICLE INFO ABSTRACT

JEL classification: We study the consequences of later marriage on subsequent life outcomes. China’s family
J12 planning policies in the early 1970s – before the One-Child Policy – regulated not only childbirth
J13 but also marriage. The recommended minimum marriage age of 25 years for men and 23
J21
years for women was effectively relaxed when the government formally introduced the One-
Keywords: Child Policy and put greater emphasis on directly controlling fertility rather than marriage.
China’s family planning policy Subsequently, we find that the marriage age, which had been increasing steadily since 1970,
Regression probability jump and kink
suddenly started to decline in the early 1980s. This policy shift provides us with an opportunity
Later marriage
to apply a regression probability jump and kink design for the purpose of identification. Using
data from the 2000 census, we establish that later-married men have fewer children and that
later-married women are more likely to participate in the labor market. We find no consistent
evidence that later marriage improves education, probably because most Chinese people marry
after completing their education.

1. Introduction

People all over the world are getting married later. According to the 2020 World’s Women Report (United Nations, Department
of Economic and Social Affairs, Statistics Division, 2020), the mean age at first marriage increased for both women and men in
all regions, with few exceptions. Globally, during the 1995–2015 period, the mean age at first marriage rose from 22.5 years to
23.3 years for women and from 25.6 years to 26.8 years for men.1 Therefore, understanding the long-term consequences of later
marriage is an important empirical question. However, it is a challenging one because the timing of marriage is an endogenous
variable associated with a wide range of socio-economic factors, such as educational attainment (Ikamari, 2005; Zha, 2019),
economic prospects (Cherlin, 1980; Oppenheimer, 1988; Bergstrom and Schoeni, 1996; Gutiérrez-Domènech, 2008; Gibson-Davis,
2009), family structure (Peter et al., 2018), and demographic changes (Yu and Xie, 2015; Porter, 2016; Bronson and Mazzocco,
2018). We contribute to the literature by establishing a causal effect of later marriage on subsequent life outcomes by exploiting

✩ We would like to thank the editor, Hongbin Li, and two anonymous referees for suggestions that significantly improved the paper. We are grateful for
the valuable comments from Yvonne Jie Chen, Rufei Guo, Sen Ma, Robert A. Moffitt, Fei Wang, Sen Xue, Zizhong Yan, and Danyan Zha. This paper has been
presented at the 2020 International Symposium on Contemporary Labor Economics. Comments from all conference participants were highly appreciated. We wish
to acknowledge the National Bureau of Statistics of China, which provided the underlying data that made this research possible. Chen also acknowledges the
financial support from the National Science Foundation of China (71903066). The two authors equally contribute to this research. Data and codes for replication
is available at https://drive.google.com/file/d/1A31bEDwGMBCpXux-NPOCrk85pCkBUXyM. All remaining errors are our own.
∗ Corresponding author.
E-mail addresses: [email protected] (Y. Chen), [email protected] (Y. Zhao).
1 https://worlds-women-2020-data-undesa.hub.arcgis.com/app/f13f7dc3b9744d5489e7ae3e87c57d71

https://doi.org/10.1016/j.jce.2022.04.005
Received 11 November 2021; Received in revised form 25 March 2022; Accepted 24 April 2022
Available online 20 May 2022
0147-5967/© 2022 Association for Comparative Economic Studies. Published by Elsevier Inc. All rights reserved.
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 1. Trends of average age at first marriage, 1970–2000.

Source: The average age at first marriage is based on our
calculation according to the 2000 census. GDP per capita
(measured in 2000 USD) data are from World Bank (2020).

changes in the late-marriage regulations as a result of a shift in China’s family planning policies, which generated dramatic kinks
(slope change) with small jumps (level change) in marriage age across birth cohorts.
The existing literature on China’s family planning policies has disproportionally focused on the fertility restrictions of the One-
Child Policy (OCP), which was formally introduced in January 1980. However, the OCP is a subset of China’s family planning
policies during a particular period. In the early 1970s, before the enforcement of the OCP, China’s family planning policies took
the form of the ‘‘later (marriage), longer (intervals), fewer (children)’’ (LLF, or wan, xi, shao) campaign (Chen and Huang, 2020).2
The ‘‘later (marriage)’’ part of the policy regulated the timing of marriage: the recommended minimum marriage age was 25 years
for men and 23 years for women. Despite imperfect enforcement, the average age at first marriage increased steadily from 23 years
in 1971 to 24.6 years in 1979 for men, and from 20.5 years to 22.8 years for women over the same period. This rising trend was
suddenly reversed in 1980 when the OCP came into effect (Fig. 1): the average age at first marriage dropped to 23.3 years for men
and 21.5 years for women in 1985.3 Our paper focuses on the shift in the late-marriage regulations and explores how they affected
the timing of first marriage and post-marriage outcomes, including fertility, education, and labor-force participation.
Using data from China’s 2000 population census and adopting a regression probability jump and kink (RPJK) strategy that
combines a regression discontinuity design with a regression kink design, we observe that (1) the relaxation of the late-marriage
requirements resulted in young couples getting married earlier, (2) men marrying later on average have fewer children, and (3)
women marrying later are more likely to participate in the labor force. We find suggestive evidence that improved intra-household
bargaining position is one potential mechanism that accounts for the causal effect of later marriage. This finding is consistent
with the literature showing that later childbirth benefits women’s labor-market performance through the channel of higher human
capital (Blackburn et al., 1993; Miller, 2011).
Our study contributes to two strands of literature. The first examines the causal effects of later marriage. Several studies use the
menarche age as the instrumental variable to examine the causality of delayed marriage on women’s schooling outcomes, social
attitudes, health, bargaining power, and political engagement (Field and Ambrus, 2008; Asadullah and Wahhaj, 2019; Hicks and
Hicks, 2019; Carpena and Jensenius, 2021). The use of the menarche-age instrument is limited to women in patriarchal societies,
who face strong social pressures to marry upon menarche. In a study of the United States, Wang and Wang (2017) employ state-
by-year variations in the legal minimum marriage age to construct an instrumental variable for marriage age. Using the 1980 U.S.
census, they find a positive effect of marriage delay on wages, with a larger effect for women, and suggest that the causal effects are
almost exclusively achieved through increased education. However, since minimum legal ages are primarily set in the teenage years,
very few Americans marry immediately upon reaching the minimum marriage age. We propose a novel approach to identifying the
causal effect of later marriage on Chinese youths in their twenties (rather than teenagers)—an age period that is closer to the world
average age for first marriage (United Nations, Department of Economic and Social Affairs, Statistics Division, 2020).
The second strand of literature investigates the effect of China’s family planning policies before the OCP. We acknowledge that
our first marginal contribution (identifying the causal effect of later marriage) faces an external validity issue because the policy
and economic environment of China is special in many aspects. However, highlighting the marital restriction during the LLF period

2 Appendix Figure A1 presents the time series of the total fertility rate in China. China’s fertility rate fell dramatically from 5.7 births per woman in 1969

to 2.8 in 1979 before the OCP came into effect. The fertility rate then declined marginally to 2.5 births per woman one decade after the OCP’s enforcement.
3 We will explain in detail in Section 2 why the switch from the LLF policy to the OCP effectively allowed people to get married earlier.

714
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

and its subsequent relaxation in the OCP period is sufficiently important on its own because it helps to understand the historical
transition of marriage patterns in China. Most existing studies focus on the effect of the OCP on various outcomes,4 while the
studies that focus on pre-OCP policies (including the LLF policy) are much fewer. Scharping (2003) and Whyte et al. (2015) stress
the importance of the LLF policy in reducing the fertility rate. Several studies focus on the consequences of the ‘‘fewer (children)’’
part of the ‘‘later, longer, fewer’’ policy bundle.5 However, the LLF policy imposed restrictions not only on the number of children
but also on the timing of marriage. Our paper is among the few studies that focus on the marriage restriction component of China’s
family planning policies.6
The remainder of the paper is organized as follows. Section 2 introduces the background of China’s family planning policies and
explains why the switch from the LLF policy to the OCP generated a kink in the marriage age. Section 3 describes our empirical
strategy, data, and main variables. Section 4 presents the main results and explores the potential mechanisms of later marriage.
Section 5 provides a conclusion.

2. Background

2.1. The ‘‘later marriage’’ policy of the 1970s

The history of China’s family planning policies dates back to December 1962 with the release of Document No. [62]698,
‘‘Instructions on Seriously Advocating Family Planning’’ (guanyu renzhen tichang jihua shengyu de zhishi). However, the outbreak
of the Cultural Revolution in 1966 interrupted the course of family planning. In the early 1970s, the Chinese government resumed
family planning and released Document No. [71]51, ‘‘Report on Better Implementing Family Planning Policy’’ (guanyu zuohao jihua
shengyu gongzuo de baogao). One important policy in the early 1970s was the ‘‘later, longer, fewer’’ policy. Our study is mostly
concerned with the ‘‘later’’ component, which encouraged marriage at a later age—23 years for women and 25 years for men.7 , 8
Although the LLF policy was technically voluntary and its enforcement during the 1970s was much more lenient than that of
the OCP (Zhang, 2017), it did not rely simply on persuasion or voluntary compliance (Whyte et al., 2015). First, the policy was
designed to meet the formal birth-planning targets codified in the Fourth Five-Year Plan (1971–1975) and the Fifth Five-Year Plan
(1976–1980).9 Second, to implement the LLF policy, provincial leaders established birth-planning offices and mid-level coordinating
committees, which translated central government guidelines into provincial- and local-level targets and managed the daily activities
of local birth-planning cadres (Babiarz et al., 2018). In 1973 and 1974, municipalities spearheaded a system of monthly inspections
and half-yearly adjustments to check the progress towards targets for late marriage and fertility reduction (Scharping, 2003). Third,
local-level cadres had strong career-advancement incentives to implement the LLF policy. During the intensified population control
efforts of the 1970s, the ‘‘late-marriage rate’’ emerged as one indicator for evaluating the performance of local birth-planning
units (Tien, 1983).10
Specific regulations were in place to enforce the ‘‘later marriage’’ policy. For example, before they could complete marriage
registration, the groom and bride had to obtain a letter of introduction from the heads of their work units, which gave them
effective permission to marry (Tien, 1983). The permission was often withheld for people younger than the recommended minimum
marriage age according to the ‘‘later marriage’’ policy (Coale et al., 1991). In addition, violations of the LLF policy were associated
with punishments in the form of more work assignments, public condemnation, and restricted food rations, medical care, and other
public services (Greenhalgh and Li, 1993). There were also benefits to compliance. Regularly documented rewards included the
reduction (or waiving) of fees for sterilization operations, priority for unmarried people in employment and education services,
priority in housing arrangements for couples who married later, and paid vacation after a sterilization operation (Chen and Huang,
2020).

4 Previously studied impacts include demographic structure (Ebenstein, 2010), education (Qin et al., 2017), female empowerment (Huang et al., 2021),

non-cognitive skills (Cameron et al., 2013), and eye health (Zhao and Zhou, 2018).
5 Babiarz et al. (2018) and Chen and Huang (2020) exploit provincial variations in the enforcement of the LLF policy to examine its effect on fertility. Chen

and Fang (2021) further examines the long-term consequences on the well-being of elderly parents forty years after the initial implementation of the LLF policy.
6 Babiarz et al. (2018) study the effect of the LLF policy on women’s marriage age. The focus of our study is related to but different from the ‘‘later

(marriage)’’ component of the LLF policy. We highlight the de facto relaxation of late-marriage requirements when family planning policies transitioned from
the LLF policy to the OCP.
7 The threshold for late marriage (23 years old for women and 25 years old for men) was homogeneous across rural China. The situation in urban China

was more complex. Some provinces set higher standards for late marriage—e.g., 25 years old for women and 28 years old for men in Shanxi, Heilongjiang,
Anhui, Fujian, Hunan, Guangdong, and Guizhou (Chen and Huang, 2020). However, this disparity is unlikely to affect our main results because China was
predominantly rural throughout the 1970s.
8 ‘‘Longer’’ refers to a birth-planning rule of waiting more than three years between births. ‘‘Fewer’’ means that one couple could have two children at most.
9 The Fourth Five-Year Plan set targets for a natural population growth rate of 1 percent in cities and 1.5 percent in the countryside. The Fifth Five-Year

Plan further lowered targets for the population growth rate to 0.6 percent for cities and 1 percent for the countryside.
10 The ‘‘late-marriage rate’’ refers to the percentage of women and men married during a calendar year who meet the late-marriage requirements. The higher

the rate, the more the local unit was commended for its achievement (Tien, 1983).

715
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 2. Number of marriages each year, 1970–1990.

Source: Our calculation based on the 2000 census.

2.2. The policy shift from LLF to the OCP and the relaxation of the late-marriage requirements

The enforcement of the ‘‘later marriage’’ policy was effectively relaxed after China’s family planning policies switched from
the LLF policy to the OCP, a much more coercive version of the ‘‘fewer children’’ policy. The OCP was officially launched in
January 1980.11 The OCP permitted only one child per couple in urban China. A second child was sometimes allowed in rural
China depending on the local policy: some counties allowed all couples to have a second child, while others permitted a second
birth only if the first-born child was female. Exceeding the quota resulted in a heavy penalty, including a one-time fine of between
one and five times the person’s annual income (Ebenstein, 2010).
Most relevant to our study is the fact that the late-marriage requirement was no longer emphasized when the family planning
policies switched from the LLF policy to the OCP. The de facto relaxation of late-marriage regulations took place for several reasons.
First and foremost, the government put greater emphasis on the birth rate than on other indirect measures. In July 1978, at the
first conference of the national Family Planning Leading Group, Chen Muhua, the newly appointed director of the group pointed
out that the crux of the LLF policy was the ‘‘fewer children’’ policy and that more efforts should be made in this direction (Peng,
1997). An official document released in September 1980, ‘‘Explanation on ‘‘the Marriage Law of the People’s Republic of China
(Revised Draft)’’, also clarified the relative importance of the birth rate over late marriage: The goal of family planning was to curb
population growth; even if people married very late, they could still have many children. Second, the new marriage law, which was
passed in 1980 and came into effect in 1981, facilitated early marriage for young couples. The new marriage law raised the legal
minimum marriage age from 20 years to 22 years for men and from 18 years to 20 years for women. Although the legal minimum
marriage age rose, it remained lower than that of the late-marriage requirements. A document (‘‘Report on Family Planning Work’’)
released by the State Family Planning Commission in March 1984 further de-emphasized late-marriage requirements: young people
reaching the legal minimum marriage age but not the age required by the ‘‘later marriage’’ policy were to be allowed to marry if
they insisted on doing so.
Following the relaxation of late-marriage requirements, we observe an immediate boom in marriages in 1980 (Fig. 2)12 and a
subsequent decline in the average age at first marriage in the early 1980s (Fig. 1). The average age at first marriage, which had
risen from 23.4 years in 1973 to 24.6 years in 1979 for men and from 21.1 years to 22.8 years for women, began to decline in
1980 when people under the age required by the ‘‘later marriage’’ policy were suddenly allowed to marry. The boom in marriages
in 1980 also partially explains the upturn in the total fertility rate in 1981 and 1982 (Appendix Figure A1), even after the coercive
OCP became effective (Coale et al., 1991).

11 Although the exact timing of the OCP’s implementation is still debated, historical records indicate that it took place between late 1979 and early 1980.

In August 1979, the Party made public a new birth-planning target that would ensure a 0.5 percent decline of the population growth rate by 1985 and a
further decline to zero growth by the turn of the century. The Party also suggested that society should encourage couples to have only one child to meet
such an ambitious target. In December 1979, at a new conference for the directors of regional birth-planning bureaus, the central government demanded that
emphasis be put on propagating the one-child family model. One-child certificates offering a variety of benefits to compliant couples started to be widely issued
in November 1979, and fines for defiant couples began to be imposed nationwide in January 1980 (Qin et al., 2017). Finally, in a circular to all Party members
dated January 1980, the Secretariat of the Central Committee sanctioned such efforts by calling for legal, economic, and administrative measures in favor of
one-child families (Scharping, 2003).
12 Other events may also have led to this marriage boom (e.g., the marriage of the returned sent-down youths). We address those potential confounders in

Section 4.3.

716
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 3. Effect of the policy shift in late-marriage requirements on the probability of late marriage. Note: The circles in the figure represent the proportion of
individuals in a birth cohort who are classified as ‘‘late marriage’’, and the vertical lines refer to the threshold birth cohort.

3. Empirical strategy

We define the ‘‘late marriage’’ dummy as whether a man (woman) gets married at or later than the age of 25 (23) years. The
unexpected relaxation of the late-marriage requirements provides a good natural experiment for an RPJK design. The first birth
cohort influenced by the policy shift is the one that reached the minimum marriage age required by the ‘‘later marriage’’ policy in
January 1980 (i.e., men born in January 1955 and women born in January 1957). We define those cohorts as threshold (or cutoff)
cohorts and normalize them to 0 in the RPJK design. Fig. 3 presents the proportion of individuals complying with late-marriage
requirements for each birth cohort and finds a sharp kink for both men and women. The following example sheds light on the origin
of the kink. Two men born in 1955 and 1956, respectively, did not meet the late-marriage requirements before 1980. If they both
married immediately in 1980, once the government relaxed the regulation, they would be aged 25 years and 24 years, respectively,
at the time of their first marriage. Therefore, the man born in 1955 is more likely to be labeled as entering into a ‘‘late marriage’’
according to our definition.

3.1. RPJK framework

Identifying the causal effect of later marriage is no easy task. Marriage is one of the most important life decisions, which is made
carefully by taking into account various factors, including but not limited to economic conditions, educational attainment, career
aspirations, and attitude towards the quality and quantity of children. For example, if we observe a positive association between
later marriage and female labor-force participation, it is possible that causality works in the opposite direction—it is not that later
marriage encourages women to join the labor force but that women who are more eager to pursue a career choose to get married
later. To overcome the endogeneity challenge, we adopt an RPJK strategy that exploits a unique policy shift in China.
Dong (2018) first proposed the RPJK design, which combines a regression discontinuity design (RDD, a level change) with a
regression kink design (RKD, a slope change). The RPJK design applies to a binary treatment and is valid regardless of whether
there is a jump, a kink, or both in the treatment probability. Therefore, this design is suitable for our setting when institutional
changes can lead to either a change in level or a change in slope.
Let 𝑌𝑖 denote the outcome variable of individual 𝑖 and 𝑇𝑖 denote the treatment variable indicating whether individual 𝑖 marries
at or after 25 (23) years of age for men (women). 𝑌𝑖1 and 𝑌𝑖0 are the potential outcomes of individual 𝑖 marrying late or not,
respectively. Thus, the causal effect of later marriage is

𝛾 = 𝐸[𝑌𝑖1 − 𝑌𝑖0 ]. (1)

For any function 𝐻(𝑣), let 𝐻+ ≡ lim𝑣↓𝑣0 𝐻(𝑣) and 𝐻− ≡ lim𝑣↑𝑣0 𝐻(𝑣) be the right and left limits at 𝑣0 (i.e., the cutoff), respectively;
let 𝐻+′ ≡ lim𝑣↓𝑣0 𝜕𝐻(𝑣)∕𝜕𝑣 and 𝐻−′ ≡ lim𝑣↑𝑣0 𝜕𝐻(𝑣)∕𝜕𝑣 be the right and left limits of derivatives at 𝑣0 , respectively, when they exist.
With the relaxation of the late-marriage requirements, it is possible that (1) the probability of late marriage is discontinuous at
the cutoff point 𝑣0 , that is, 𝑇+ |(𝑣𝑖 = 𝑣0 ) ≠ 𝑇− |(𝑣𝑖 = 𝑣0 ), and that (2) the slope in the probability also changes discontinuously at the
cutoff point 𝑣0 , that is, 𝑇+′ |(𝑣𝑖 = 𝑣0 ) ≠ 𝑇−′ |(𝑣𝑖 = 𝑣0 ). Under the smoothness and monotonicity assumptions required to identify a local
average treatment effect, Dong (2018) shows that the RPJK estimand identifies the treatment effect at the cutoff point 𝑣0 :
′ ′
𝑌𝑖+ − 𝑌𝑖− + 𝑤𝑛 (𝑌𝑖 + − 𝑌𝑖 − )
𝛾 𝑅𝑃 𝐽 𝐾 = 𝑙𝑖𝑚 ′+ ′−
, (2)
𝑤𝑛 →0 𝑇𝑖+ − 𝑇𝑖− + 𝑤𝑛 (𝑇𝑖 − 𝑇𝑖 )
where 𝑤𝑛 is any sequence of nonzero weights such that lim𝑛→∞ 𝑤𝑛 = 0.

717
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Table 1
Summary statistics.
Panel (a) Men Panel (b) Women
Bandwidths (months) 12 24 48 12 24 48

Marriage
Age at First Marriage 24.48 24.44 24.33 22.44 22.42 22.31
(3.24) (3.27) (3.31) (2.76) (2.80) (2.89)
Late Marriage = 1 0.49 0.47 0.45 0.52 0.50 0.47
(0.50) (0.50) (0.50) (0.50) (0.50) (0.50)
Observations 166,388 318,364 611,592 164,016 312,282 582,540
Age Difference between Husband 2.07 2.08 2.12 -1.98 -1.99 -2.00
and Wife (3.28) (3.28) (3.24) (3.06) (3.09) (3.14)
Observations 153,586 293,762 563,484 150,132 285,761 532,203

Fertility
The Number of Children 2.06 2.07 2.08 2.06 2.06 2.08
(1.01) (1.01) (1.01) (1.01) (1.01) (1.02)
Observations 152,923 292,164 556,246 164,016 312,282 582,540

Education
Years of Schooling 8.50 8.48 8.48 7.12 7.11 7.09
(3.05) (3.06) (3.06) (3.64) (3.63) (3.64)
Observations 166,116 317,832 610,606 163,695 311,650 581,348

Performance in Labor Market

Labor-Force Participation 0.97 0.97 0.97 0.86 0.85 0.84
(0.16) (0.16) (0.17) (0.35) (0.35) (0.36)
Observations 166,388 318,364 611,592 164,016 312,282 582,540

Note: This table presents the means and standard deviations (in parentheses) of key variables for individuals
within the bandwidths listed at the top of each column. The number of observations varies by variable within
a column because of missing values.

The smoothness and monotonicity assumptions are two necessary conditions of the RPJK design (Dong, 2018). The smoothness
assumption is that individuals cannot precisely manipulate the running variable (i.e., the timing of childbirth) to sort around the
cutoff point. The smoothness assumption is intuitively valid. Individuals and their parents cannot predict the ‘‘later marriage’’ policy
and the subsequent relaxation and, thus, would not manipulate the timing of childbirth based on the policy. In addition, cesarean
sections and ultrasounds were not widely available across China before 1970 (Li and Zheng, 2009; Chen et al., 2013), making it
difficult to manipulate the timing of childbirth even if the parents wanted to. The monotonicity assumption in our setting states
that no one chose to get married early when the late-marriage requirements were stringent but decided to get married late after the
relaxation. We will provide suggestive evidence in favor of this assumption in Section 4.3.
The RPJK estimand, 𝛾 𝑅𝑃 𝐽 𝐾 , was identified by a local two-stage least squares (2SLS) estimation using both 𝐷𝑖 (≡ 𝐼[𝑣𝑖 ≥ 𝑣0 ]) and
𝐷𝑖 ⋅(𝑣𝑖 −𝑣0 ) as the instrumental variables for 𝑇𝑖 . The local 2SLS estimation provides weights reflecting the relative strength of the two
instrumental variables (IVs) that have the required properties. Using the local 2SLS estimation requires choosing a kernel function,
a bandwidth, and an order of the polynomial. For the choice of kernel function and order of the polynomial, we follow Dong (2018)
and Card et al. (2015b) and use a uniform kernel function13 and local linear regression.14 Regarding the choice of the optimal
bandwidth, there is active econometrics literature in the RDD and RKD settings (Fan et al., 1996; Lee and Lemieux, 2010; Imbens
and Kalyanaraman, 2012; Calonico et al., 2014; Card et al., 2015a). We use the optimal data-driven bandwidth chosen by Calonico
et al. (2014) (CCT optimal data-driven bandwidth) because this selection procedure is valid even when the running variable 𝑣𝑖 is
discontinuous. We also conduct analyses with all possible bandwidths in one-month increments from 12 months to 48 months to
check the sensitivity of our results.
More specifically, we estimate the following first-stage regression for the local 2SLS estimation:

𝑇𝑖,𝑚,𝑝 = 𝛿1 + 𝛼1 𝐷𝑖 + 𝛽1 𝐷𝑖 ⋅ (𝑣𝑖 − 𝑣0 ) + 𝜏1 (𝑣𝑖 − 𝑣0 ) + 𝜔𝑝 + 𝜒𝑚 + 𝜌′ 𝐗𝑖 + 𝜖𝑖,𝑚,𝑝 if |𝑣𝑖 − 𝑣0 | ≤ ℎ, (3)

for the subsample within a narrow bandwidth ℎ surrounding the cutoff point 𝑣0 . 𝑇𝑖,𝑚,𝑝 is the treatment variable (late marriage) for
individual 𝑖 born in month 𝑚 in province 𝑝. 𝛼1 and 𝛽1 are the estimated jump and kink, respectively, in the treatment variable at the
cutoff point. We control for birth-month fixed effects (𝜒𝑚 ) to address the possibility that the season of birth may be associated with
post-marriage outcomes (Angrist and Krueger, 1992; Buckles and Hungerman, 2013). We also control for birth-province fixed effects
𝜔𝑝 . 𝐗𝑖 refers to a vector of individual controls including a dummy variable for whether an individual held an urban hukou at the

13 We use a triangular kernel function as a robustness check in Tables 2 and 3.

14 Local linear regression is a suitable order of the polynomial in the RPJK setting. Gelman and Imbens (2019) argued that controlling global high-order
polynomials in regression discontinuity analysis is a flawed approach and recommended that researchers use estimators based on local linear or quadratic
polynomials. However, Card et al. (2015b) found that the local quadratic estimators for the RKD are generally quite noisy, leaving local linear regression as the
most viable option for us.

718
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Table 2
The policy shift in late-marriage requirements and the timing of marriage.
ESTIMATE VARIABLE
Late Marriage (Percentage Point, %)
(1) (2) (3)
Panel (a) Men
Estimated Jump 1.496* 0.946** 2.044*
(0.836) (0.457) (1.036)
Estimated Kink −0.481*** −0.465*** −0.473***
(0.065) (0.029) (0.096)
Bandwidth 25 25 26
Observations 331,515 331,515 343,740
Panel (b) Women
Estimated Jump 0.555 −0.259 0.930
(0.801) (0.477) (0.699)
Estimated Kink −0.469*** −0.478*** −0.473***
(0.060) (0.027) (0.047)
Bandwidth 25 25 38
Observations 323,751 323,751 473,485
Kernel Uniform Uniform Triangular
Covariates N Y Y

Note: Panels (a) and (b) present the estimated effect of the policy shift in late-marriage
requirements on the probability of later marriage for men and women, respectively. Each cell
shows the estimated jump and kink at the cutoff point. We use local linear regressions and the
CCT optimal data-driven bandwidth. The first two columns use a uniform kernel function, and
the third column uses an alternative triangular kernel function as the robustness check. Standard
errors in parentheses are clustered at the level of the running variable. * significant at 10 percent;
** significant at 5 percent; *** significant at 1 percent. Columns (2) and (3) additionally control
for the urban hukou, exposure to the LLF policy, exposure to the OCP, dummies for the province
of birth, and dummies for the month of birth.

Table 3
Effects of later marriage on subsequent outcomes.
ESTIMATE VARIABLES
Children Ever Born Years of Schooling Labor-Force Participation
(1) (2) (3) (4) (5) (6)
Panel (a) Men
Late Marriage −0.553*** −0.508*** 1.024*** 0.932*** 0.025 0.027*
(0.103) (0.095) (0.203) (0.197) (0.018) (0.016)
Bandwidth 19 26 29 36 22 34
Observations 234,201 315,365 381,010 475,617 293,654 452,358
Panel (b) Women
Late Marriage 0.252 −0.165 0.390* −0.144 0.262*** 0.254***
(0.190) (0.433) (0.233) (0.315) (0.024) (0.044)
Bandwidth 17 15 27 34 31 24
Observations 228,090 202,737 346,754 425,368 391,674 312,282
Kernel Uniform Triangular Uniform Triangular Uniform Triangular
Covariates Y Y Y Y Y Y

Note: Panels (a) and (b) present the estimated effect of later marriage on subsequent outcomes for men and women,
respectively. We show the results for the number of children ever born, years of schooling, and labor-force participation.
We use local linear regressions and the CCT optimal data-driven bandwidth. The odd (even) columns use a uniform
(triangular) kernel function. Standard errors in parentheses are clustered at the level of the running variable. * significant
at 10 percent; ** significant at 5 percent; *** significant at 1 percent. Covariates include the urban hukou, exposure to
the LLF policy, exposure to the OCP, dummies for the province of birth, and dummies for the month of birth.

time of the survey, exposure to the LLF policy, and exposure to the OCP (defined in the following subsection).15 China’s household
registration assigns people an urban or rural hukou status, and urban hukou holders enjoy much better social welfare than rural
hukou holders. Exposure to the LLF policy and the OCP controls the confounding effect of fertility reduction caused by the LLF
policy and the OCP, respectively. 𝜖𝑖,𝑞,𝑝 is the unobserved error term, and standard errors are clustered at the year–month-of-birth
level (i.e., the level of the running variable 𝑣𝑖 ), as suggested by Lee and Card (2008).

15 Including covariates is not necessary to obtain consistent estimates of the treatment effect for the RDD and RKD settings (Lee and Lemieux, 2010; Card

et al., 2015b). However, adding covariates in regressions could reduce sampling variability (Lee and Lemieux, 2010) and address the estimation bias induced
by some confounding non-linearity between the outcome variable and the running variable in the RKD setting (Ando, 2017).

719
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

The second-stage regression is:

𝑌𝑖,𝑚,𝑝 = 𝛿2 + 𝛽2 𝑇̂𝑖,𝑚,𝑝 + 𝜏2 (𝑣𝑖 − 𝑣0 ) + 𝜓𝑝 + 𝜂𝑚 + 𝜁 ′ 𝑋𝑖 + 𝜀𝑖,𝑚,𝑝 if |𝑣𝑖 − 𝑣0 | ≤ ℎ. (4)

We instrument 𝑇𝑖,𝑚,𝑝 with 𝐷𝑖 and 𝐷𝑖 ⋅ (𝑣𝑖 − 𝑣0 ). The coefficient of interest, 𝛽2 , provides an estimate of 𝛾 𝑅𝑃 𝐽 𝐾 and measures the effect
of later marriage (𝑇𝑖,𝑚,𝑝 ) on the outcome variables (𝑌𝑖,𝑚,𝑝 ).16

3.2. Are fertility restrictions a threat to the RPJK design?

Because the late-marriage regulations were part of the bundle of China’s family planning policies in the 1970s, a natural question
is whether other policies (especially restrictions on childbirth) would contaminate our results. We believe that this is unlikely. The
RPJK design estimates the local effect for the population that was about to marry in 1980. Because premarital pregnancy was rare
in China at the time – with a rate of approximately 4 percent in the early 1980s (Wang and Yang, 1996) – the young couples should
have expected their first child no earlier than 1981. Therefore, cohorts around the cutoff 𝑣0 were exposed to the same version of
family planning, that is, the OCP.
Nevertheless, we control for exposure to fertility restrictions in our baseline specification. Following Chen and Huang (2020),
we define exposure to the LLF policy and exposure to the OCP as follows:
∑
49
LLF𝑝,𝑐 = AFR𝑝,1969 (𝑎) × 𝐼[𝑇𝑝 ≤ 𝑐 + 𝑎 ≤ 1979]
𝑎=15
(5)
∑49
OCP𝑝,𝑐 = AFR𝑝,1979 (𝑎) × 𝐼[1980 ≤ 𝑐 + 𝑎],
𝑎=15

where LLF𝑝,𝑐 and OCP𝑝,𝑐 define exposure to the two programs, respectively, for cohorts born in year 𝑐 in province 𝑝. AFR𝑝,𝑡 (𝑎) refers
to the age-specific fertility rate at age 𝑎 in province 𝑝 before the implementation of the policy (1969 for the LLF policy and 1979
for the OCP).17 , 18 𝐼[start year𝑝 ≤ 𝑐 + 𝑎 ≤ end year𝑝 ] is an indicator variable equal to 1 if the policy is effective in province 𝑝 when
cohort 𝑐 reaches age 𝑎 and to 0 otherwise. The LLF policy in province 𝑝 started in 𝑇𝑝 —the year when the Family Planning Leading
Group was established in province 𝑝 (Chen and Huang, 2020)—and ended in 1979. The OCP started in 1980. The measurement can
be interpreted as the number of children an individual would have given birth to under the influence of the policy (LLF/OCP) when
retaining the pre-policy fertility pattern at the province level.

3.3. Data and sample

We use a one-percent sample from the 2000 wave of the Chinese census obtained from IPUMS International (Minnesota
Population Center, 2020). We exclude individuals who are non-Han or whose spouses are non-Han because most ethnic minorities
are subject to different family planning policies. Never-married individuals are also excluded.19 We drop the top and bottom 0.5
percent of the sample according to the age at first marriage to alleviate the influence of outliers. We also drop the Tibet sample
because of the lack of information about provincial family planning policies. The final sample consists of approximately 6.5 million
observations. We do not impose any further restrictions and let the optimal data-driven bandwidth decide the final sample of the
local 2SLS estimation.
The 2000 census is the most suitable wave for this study because it is the first population census to include the age at first
marriage. The age range of our cohorts of interest in 2000 is also suitable for our study. Because we focus on the results with
a bandwidth of 12 months to 48 months, the birth cohorts of interest consist of men born between January 1951 and January
1959 and women born between January 1953 and January 1961. Most of those cohorts should have completed their fertility and
education by 2000. Earlier censuses, including the 1982 and 1990 censuses, do not satisfy this condition. Additionally, most cohorts
did not reach the legal retirement age by 2000, which is 60 for men, 50 for female workers, and 55 for female cadres. The oldest
female cohort (January 1953) was 47 years old by 2000 and may have retired in more recent censuses. We also use a one-percent
sample from the 1990 census and a 20 percent sample from the 2005 one-percent-inter-decennial census for supplemental analysis.

16 IV estimator is a Wald estimator that equals the ratio of the reduced-form coefficient to the first-stage coefficient. In our baseline specification (Eqs. (3)

and (4)), we use the same bandwidth for both the first-stage and the reduced-form estimations. An alternative approach (such as that of Jiang et al. (2020)) is
to allow for different bandwidths for the two estimations. Appendix Table A1 reports the corresponding results, which are fairly similar to our main results in
Table 3.
17 We drew the data on province- and age-specific fertility rates from Coale and Li (1987).
18 The definition of the exposure variable is somewhat more complex for men because the calculation of the fertility rate mechanically uses the female

sample. For men living with their wives, we can define exposure according to their wives’ birth cohorts. For other men, we take the stylized fact that the
average husband–wife age gap is approximately two years in China and define their policy exposure according to their own age minus two.
19 There may be concern that the two restrictions (ethnic-Han and married samples) raise the issue of sample selection. We first examine the demographic

characteristics around the cutoff point and find no significant change in the proportion of ethnic minorities. Regarding the ever-married sample restriction, we
highlight that a vast majority of Chinese people eventually got married. For people aged 45 or above in the 2000 census, only 1.97 percent never married (3.72
percent for men and 0.2 percent for women). To further evaluate the impact of sample selection of the ever-married, we perform the following exercise. Whereas
the IV regression with late marriage as the endogenous variable certainly requires information on the date of the first marriage, the reduced-form specification
does not. Therefore, we can compare the reduced-form results with and without the ever-married restriction (Appendix Table A2), and we find that the sample
restriction yields negligible distinctions in the estimated jumps and kinks.

720
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

3.4. Variables

We define the treatment variable (late marriage) as 1 if a man (woman) marries at or after 25 (23) years of age. We focus on three
outcome variables: the number of children, years of schooling, and labor-force participation. Information on the number of children
is only available for women aged between 15 and 50 years in 2000. We construct this variable for men according to their wives’
information. If the wife’s information is unavailable for any reason (e.g., the husband and wife did not live together at the time of
the survey), we code the number of children as missing.20 The 2000 census does not directly survey years of schooling. Therefore,
we code this variable according to the highest level of education received and the school-completion status, following Chen et al.
(2020). If individuals completed primary school, we assume that they received six years of education. If they dropped out of or
were attending primary school, we coded the number to three. Higher-level schooling years were coded similarly. The final outcome
variable of interest is labor-force participation. We consider that a person was participating in the labor force if he/she had worked
more than one hour in the previous week or was seeking a job.
Table 1 presents the summary statistics of key variables for the sample with different bandwidths (12, 24, and 48 months).
Approximately half of the sample is labeled as ‘‘late marriage’’ according to our criteria, and the husband is on average two years
older than his wife. Individuals have two children on average. Men have 8.5 years of schooling on average, compared with 7.1 years
for women. The labor-market participation rate in China is especially high, probably because many people work in the agricultural
sector. At the time of the survey, men almost universally (over 97 percent) participated in the labor market. Women’s rate of
labor-market participation (about 85 percent) was lower than men’s but high by international standards.
Fig. 3 shows that cohorts closer to the cutoff were more likely to marry late. The second row of Table 1 confirms this fact: for the
male (female) sample, the rate of late marriage declined from 0.49 (0.52) for the 12-month window to 0.47 (0.50) for the 24-month
window, and further to 0.45 (0.47) for the 48-month window. One important observation from Table 1 is that the narrower the
window, the more likely people were to marry late, marry someone of a similar age, have fewer children, receive more education,
and participate in the labor market. This observation provides suggestive evidence of the potential benefits of later marriage.

4. Empirical results

4.1. (First stage) how did the relaxation of the late-marriage requirements affect the timing of marriage?

Fig. 3 presents the probability of late marriage for men and women against birth cohorts (the running variable). The vertical lines
in each panel denote the cutoff birth cohort (January 1955 for men and January 1957 for women). The figure shows a noticeable
kink at the cutoff cohort, suggesting that after the de facto relaxation of the late-marriage requirements, young couples tended to
get married earlier.
Table 2 reports the first-stage estimates for men and women, respectively.21 Column (1) displays strong and statistically significant
kinks. The −0.481 coefficient means that for a cohort that is one year younger than the threshold cohort, the chances of late marriage
are 5.8 percentage points (−0.481 × 12) lower than in a counterfactual scenario of a continued rising trend of late marriage without
the policy shift. The probability of late marriage shows no significant jump for women and a marginally significant jump of 1.5
percentage points for men. Controlling for a wide range of covariates (including urban hukou, exposure to the LLF policy, exposure
to the OCP, dummies for province of birth, and dummies for month of birth) in column (2) barely alters the coefficient, providing
suggestive evidence of the exogeneity of the policy shift. Column (3) adopts a triangular kernel function (instead of a uniform kernel
function) as a robustness check and finds that using the alternative kernel function has a limited impact on our first-stage estimates.

4.2. (Second stage) the effect of later marriage on post-marriage outcomes

The left panels of Figs. 4A and 4B present post-marriage outcomes (including fertility, education, and labor-force participation)
for each cohort.22 Before proceeding to the formal RPJK analysis, we first look at the raw data and check whether there are
corresponding kinks at the cutoff cohort. We observe strong kinks in men’s number of children and women’s labor-force participation.
The former is in the opposite direction compared to the kink in the first stage, but the latter is in the same direction, suggesting
that later marriage has a negative effect on men’s number of children and a positive effect on women’s labor-force participation.
Table 3 shows the second-stage estimation for men and women using a local 2SLS approach. The odd columns use the uniform
kernel function, and the even columns use the triangular kernel function. Again, our results are not sensitive to the choice of kernel
functions, so we choose the uniform kernel function as our baseline setting. Consistent with the patterns in the left panels of Figs. 4A
and 4B, the regression results reveal a negative effect of later marriage on men’s number of children (−0.55, column [1] in panel
[a]) and a positive effect on women’s labor-force participation (26 percentage points, column [5] in panel [b]). The CCT optimal

20 Concerns might arise when coding the missing values differently for men and women. Therefore, as a robustness check, we re-run our estimations using a

subsample of co-residing couples. Their results are similar and are available upon request.
21 Note that the numbers of observations in the RPJK tables do not necessarily match those in the summary table (Table 1) because we use data-driven

optimal bandwidth, which can vary across specifications.

22 We are aware that people receive most of their education before marriage. Still, there is room for educational improvement after marriage. Our calculation

using the 2010 wave of the China Family Panel Studies (CFPS) suggests that among all post-senior-high graduates (including associate and bachelor’s degrees),
46.6 percent of men and 39.0 percent of women obtained their degrees through adult education.

721
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

data-driven bandwidths range from 15 months to 36 months. To check whether the previous two findings are sensitive to the choice
of the optimal bandwidth, we experiment with alternative bandwidths in one-month increments from 12 months to 48 months in
the right panels of Figs. 4A and 4B. We confirm that the negative effect on men’s number of children and the positive effect on
women’s labor-force participation remain statistically significant for a reasonable range of bandwidths.
Column (3) of Table 3 shows that later marriage has a statistically strong effect on men’s education and a marginally significant
effect on women’s education. The positive effect on men’s education is robust to the choice of the optimal bandwidth (right panel
of Fig. 4A), while the positive effect on women’s education is sensitive to the choice of bandwidth (right panel of Fig. 4B) and
kernel function (column [4] in panel [b]). Therefore, we find no evidence supporting a positive effect of later marriage on women’s
education.
Because our identification mainly relies on the kinks, we use a fuzzy regression kink design instead of RPJK as a robustness
check. Specifically, the first-stage regression of the regression kink design is (Eq. (3) without the jump)

𝑇𝑖,𝑚,𝑝 = 𝛿1 + 𝛽1 𝐷𝑖 ⋅ (𝑣𝑖 − 𝑣0 ) + 𝜏1 (𝑣𝑖 − 𝑣0 ) + 𝜔𝑝 + 𝜒𝑚 + 𝜌′ 𝐗𝑖 + 𝜖𝑖,𝑚,𝑝 if |𝑣𝑖 − 𝑣0 | ≤ ℎ, (6)

and the second-stage regression only uses the kink, 𝐷𝑖 ⋅ (𝑣𝑖 − 𝑣0 ) as the instrument variable of late marriage. We obtain almost
identical results with the regression kink design, as shown in Appendix Table A3.
We further complement our analysis using the 2005 one-percent-inter-decennial census. Appendix Table A4 and Appendix Figure
A2 present the results using the 2005 mini-census data. The first three columns of Appendix Table A4 double-check our previous
findings using the 2005 mini-census. We verify that later marriage induces men to have fewer children (column [1]) and encourages
women to participate in the labor force (column [3]). The last column of Appendix Table A4 provides some weak evidence that
later marriage improves women’s monthly income conditional on labor-force participation. However, Appendix Figure A2 suggests
that the finding regarding women’s income is sensitive to the choice of bandwidth. Therefore, we reserve our judgment regarding
this finding.
To summarize, we highlight three potential effects of later marriage: (1) a reduction in men’s number of children, (2) an increase
in women’s rate of labor-force participation, and (3) an improvement in men’s educational achievement. In the next subsection, we
will show that the first two effects are robust to a wide range of checks, while the last effect (on men’s education) is less clear and
is sensitive to potential confounders. Therefore, our conclusion only concerns the first two findings.

4.3. Identification assumptions

This subsection first discusses the identification assumptions for general regression-discontinuity style specification (the smooth-
ness and monotonicity assumptions) and then explores several confounding factors that are specific to our empirical context.

Smoothness and monotonicity assumptions

The smoothness and monotonicity assumptions are the two basic assumptions of the RPJK design (Dong, 2018). The smoothness
assumption is that 𝐸[𝑌𝑖0 |𝑣𝑖 = 𝑣] and 𝐸[𝑌𝑖1 |𝑣𝑖 = 𝑣] are continuously differentiable in the neighborhood of 𝑣𝑖 = 𝑣0 . Following the
arguments in Lee and Card (2008) and McCrary (2008), this smoothness assumption in practice implies that individuals cannot
precisely manipulate the running variable (i.e., the timing of birth) to sort around the cutoff point. Because the policy shift took
place twenty years after the birth of the relevant cohorts, neither those cohorts nor their parents could manipulate the timing of
childbirth based on the expectation of the policy shift.
The monotonicity assumption (or the no-defier assumption) states that the presence of the instrument never dissuades someone
from taking the treatment. In our empirical setting, this assumption implies that no one chose to get married early when the late-
marriage requirements were stringent but decided to get married late after the relaxation. Although the monotonicity assumption
is intuitively valid in our setting, we perform a test proposed by Dong (2018), who shows that when the smoothness assumption
holds, the monotonicity assumption can be assessed according to the characteristics of the marginal compliers. More specifically,
for any pre-determined characteristics 𝑄 ∈ [𝑞min , 𝑞max ], the mean characteristics of the marginal complier should fall between 𝑞min
and 𝑞max . If only the smoothness assumption holds, the identified characteristic of the marginal complier is the weighted difference
in means or distributions between compliers and defiers and, therefore, does not necessarily fall within [𝑞min , 𝑞max ].
To identify the mean value of 𝑄 for the marginal complier, we follow Dong (2018) and run the RPJK regression below with a
24-month bandwidth23 :

𝑇𝑖,𝑚,𝑝 = 𝛿1 + 𝛼1 𝐷𝑖 + 𝛽1 𝐷𝑖 ⋅ (𝑣𝑖 − 𝑣0 ) + 𝜏1 (𝑣𝑖 − 𝑣0 ) + 𝜖𝑖,𝑚,𝑝 if |𝑣𝑖 − 𝑣0 | ≤ ℎ,

𝑄𝑖 𝑇𝑖,𝑚,𝑝 = 𝛿2 + 𝛽2 𝑇̂𝑖,𝑚,𝑝 + 𝜏2 (𝑣𝑖 − 𝑣0 ) + 𝜀𝑖,𝑚,𝑝 if |𝑣𝑖 − 𝑣0 | ≤ ℎ, (7)

where 𝑇̂𝑖,𝑚,𝑝 on the right side is instrumented with 𝐷𝑖 and 𝐷𝑖 ⋅(𝑣𝑖 −𝑣0 ) at the second stage. Note that we do not impose any covariate 𝐗𝑖
because it may coincide with the pre-determined characteristics 𝑄𝑖 . Coefficient 𝛽2 identifies the mean characteristics of the marginal
complier, and we have to check whether 𝛽̂2 falls within [𝑞min , 𝑞max ].

23 Appendix II of Dong (2018) provides more details on how such a regression can identify the characteristics of the marginal complier.

722
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 4A. Effect of later marriage on subsequent outcomes (men). Note: This figure presents the effects of later marriage on subsequent outcomes for men.
The left panels show the jumps and kinks in the number of children ever born, years of schooling, and labor-force participation, respectively. The circles in
these graphs represent unconditional sample means for each cohort. The solid lines represent the fitted values from local linear regressions along with the 95
percent confidence interval. The right panels provide the results of the second-stage estimations (along with the 95 percent confidence interval) using local linear
regressions with alternative bandwidths in one-month increments from 12 to 48 months. The vertical dash line refers to the CCT optimal data-driven bandwidth.

723
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 4B. Effect of later marriage on subsequent outcomes (women). Note: This figure presents the effects of later marriage on subsequent outcomes for women.
The left panels show the jumps and kinks in the number of children ever born, years of schooling, and labor-force participation, respectively. The circles in
these graphs represent unconditional sample means for each cohort. The solid lines represent the fitted values from local linear regressions along with the 95
percent confidence interval. The right panels provide the results of the second-stage estimations (along with the 95 percent confidence interval) using local linear
regressions with alternative bandwidths in one-month increments from 12 to 48 months. The vertical dash line refers to the CCT optimal data-driven bandwidth.

724
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Table 4
Testing the monotonicity assumption—Identifying the characteristics of marginal compliers.
Characteristics (𝑄) Urban Hukou Exposure to the LLF policy Exposure to the OCP Eastern regions Central regions Western regions
(1) (2) (3) (4) (5) (6)
Panel (a) Men 0.339 0.645 2.301 0.508 0.114 0.378
(0.065) (0.117) (0.105) (0.051) (0.065) (0.056)
The 95% Confidence Interval [0.212,0.466] [0.415,0.876] [2.100,2.506] [0.409,0.607] [−0.014,0.242] [0.268,0.488]
[𝑞min , 𝑞max ] [0,1] [0,3.031] [0.02,4.925] [0,1] [0,1] [0,1]
Panel (b) Women 0.306 0.943 2.059 0.487 0.29 0.223
(0.109) (0.337) (0.191) (0.059) (0.043) (0.051)
The 95% Confidence Interval [0.092,0.521] [0.282,1.604] [1.685,2.433] [0.372,0.602] [0.206,0.375] [0.123,0.322]
[𝑞min , 𝑞max ] [0,1] [0.135,2.054] [0.975,4.522] [0,1] [0,1] [0,1]

Note: This table reports the estimated coefficient 𝛽̂2 in Eq. (4), which can be interpreted as the estimated average characteristics of marginal compliers, using
the method proposed by Dong (2018). Standard errors in parentheses are clustered at the level of the running variable.

Table 4 performs the monotonicity test. The pre-determined characteristics we chose include all covariates controlled in our main
specification and three dummies indicating the economic region in which an individual was born.24 All twelve estimated means and a
vast majority of the 95 percent confidence interval fall within the range of [𝑞min , 𝑞max ]. To summarize, our data suggests no evidence
against the monotonicity assumption.

Non-linearity of the outcome variable

An important concern for the standard RKD is the probability of spurious effects resulting from some non-linearity in the
underlying relationship between the outcome variable and the running variable (Ando, 2017; Ganong and Jäger, 2018). This is
also a valid concern for the RPJK design because our estimation mainly relies on the kinks. Therefore, we follow Ganong and Jäger
(2018) and perform a set of permutation tests. The idea is to estimate the first-stage and reduced-form regressions using placebo
cutoff points and then verify that the largest estimated kink occurs at the true cutoff point for both the first-stage and reduced-form
estimations (Manoli and Turner, 2018). Specifically, we change the placebo cutoff point in one-month increments around the true
cutoff point and estimate the pseudo first-stage and reduced-form RPJK specification for each outcome variable, using a 24-month
bandwidth around each placebo cutoff point.
Figs. 5A and 5B show the estimated kinks of the RPJK design for each placebo cutoff. The vertical line in each graph refers to the
true cutoff point. Panel (a) shows the first-stage estimates and panels (b), (c), and (d) correspond to the reduced-form estimation
results for the number of children, years of schooling, and labor-force participation, respectively. Panels (a) in Figs. 5A and 5B
highlight the fact that the largest estimated kinks appear to occur near the true cutoff point for both men and women, providing
strong evidence as to how the alleviation of the late-marriage requirements affected the timing of marriage. The reduced-form
estimates have mixed implications, however. Panel (b) of Fig. 5A (panel [d] of Fig. 5B) suggests that the kink is largest at the
genuine cutoff point for men’s number of children and women’s labor-force participation. The finding for men’s education (panel
[c] of Fig. 5A) is less robust—the pseudo kinks are statistically significant at every placebo cutoff point that is 0–24 months younger
than the true cutoff point.

Age effect instead of cohort effect?

We set the RPJK design according to cohorts instead of age. Because we mainly use a cross-sectional survey (the 2000 census)
in the empirical analysis, we cannot distinguish between the age effect and the cohort effect. Thus, we ask whether the observed
kinks in the outcome variables of interest may be driven by the age effect rather than the cohort effect. This question is particularly
important for labor-force participation, which presents strong non-linearity across the life-cycle. We use the 1990 census and the
2005 one-percent-inter-decennial census to address this concern (Fig. 6). The same cohorts are of different ages in those two census
waves. Fig. 6 confirms that the kinks in female labor participation are a result of the cohort effect instead of the age effect: the
kinks in 1990 and 2005 occur in the same cohort as that of 2000.
Regarding the other two outcome variables (childbirth and education), the inability to separate the cohort effect from the age
effect is less important. The cutoff cohorts were in their early forties in 2000, around which age their fertility and education were
unlikely to change. Therefore, if we observe any kink in childbirth and education in the 2000 census, we can reasonably suppose
that it is generated by the cohort effect and not the age effect.

24 Chinese provinces are typically classified into three economic regions: the eastern, central, and western regions. More specifically, the eastern group includes

eleven provinces (Beijing, Tianjin, Hebei, Liaoning, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, and Hainan), the central group consists of eight
provinces (Shanxi, Jilin, Heilongjiang, Anhui, Jiangxi, Henan, Hubei, and Hunan), and the western group consists of eleven provinces (Sichuan, Guizhou, Yunnan,
Tibet, Shaanxi, Gansu, Ningxia, Qinghai, Xinjiang, Guangxi, and Inner Mongolia.

725
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 5A. Permutation tests using placebo cutoffs (men). Note: The permutation tests follow the idea of Ganong and Jäger (2018). We change the placebo cutoff
points in one-month increments around the true cutoff point and estimate the RPJK model with a 24-month bandwidth. The figures plot the estimated kinks at
each placebo cutoff point for men. Panel (a) shows the first-stage estimation results and panels (b), (c), and (d) correspond to the reduced-form results for the
number of children ever born, years of schooling, and labor-force participation, respectively.

The great famine

Even if we ascertain that it is the cohort effect that generates our results rather than the age effect, we are still concerned about
other confounders that may contaminate the cohort effect. The Great Famine (1959–1961) is one such example. The threshold
cohorts (January 1955 for men and January 1957 for women) were in their early childhood when the famine broke out. The famine
killed 16.5 to 45 million individuals, most of whom lived in rural areas (Meng et al., 2015). Appendix B addresses the issue of the
Great Famine in three ways: (1) we claim ex ante that if the famine is the main driving force of the cohort effect, the kink should
appear at different positions; (2) we control for the local intensity of the famine using the method proposed by Meng et al. (2015);
and (3) we run robustness checks for the provinces that were less exposed to the famine. Appendix B yields similar implications
to the previous robustness checks: the negative effect of later marriage on men’s number of children and the positive effect on
women’s labor-force participation are robust to the inclusion of the Great Famine, whereas the positive effect on men’s education
is less robust.

The send-down movement

In the Background section, we explained why late-marriage requirements were relaxed when the family planning policies
transitioned from the LLF policy to the OCP. We affirmed that the policy shift lead to a marriage boom in 1980 (Fig. 2). However,
the send-down movement could be another reason for this boom. From 1968 to the late 1970s, the send-down movement mandated
about 16 million urban youth to temporarily resettle in rural areas (Chen et al., 2020). Most SDYs returned to their urban homes
after the discontinuation of the movement in September 1980. Many of them got married shortly after leaving the countryside.
Therefore, a large group of sent-down youths (SDYs) got married in 1980 for historical reasons. We use the 2010 wave of the China
Family Panel Studies (CFPS) to illustrate this point. In the 2010 CFPS, our calculation suggests that the number of SDY marriages
in 1980 is more than twice as high as in 1979 for women and approximately 1.3 times as high for men.
However, we contend that the send-down movement is not a big concern for our empirical strategy. First, despite their large
number, SDYs only accounted for a small portion of the whole population because the vast majority of China was rural in the

726
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 5B. Permutation tests using placebo cutoffs (women). Note: The permutation tests follow the idea of Ganong and Jäger (2018). We change the placebo
cutoff points in one-month increments around the true cutoff point and estimate the RPJK model with a 24-month bandwidth. The figures plot the estimated
kinks at each placebo cutoff point for women. Panel (a) shows the first-stage estimation results and panels (b), (c), and (d) correspond to the reduced-form
results for the number of children ever born, years of schooling, and labor-force participation, respectively.

Fig. 6. Effect of later marriage on female labor-force participation in different years. Note: This figure shows the jumps and kinks in female labor-force
participation using the 1990 census and the 2005 one-percent-inter-decennial census. The circles in the figure represent unconditional sample means of female
labor-force participation for each cohort. The solid lines represent the fitted values from local linear regressions along with the 95 percent confidence interval.

727
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

1960s and 1970s. Our calculation based on the CFPS suggests that even among the most affected cohorts,25 only 7 percent were
SDYs. Second, as a conservative robustness check, we re-run our estimations using only the rural sample not directly involved in
the send-down movement and report the results in Appendix Table A5. For the number of childbirths and the years of education,
the coefficients using only the rural sample are quite similar to those using the whole sample (Table 3). Although the coefficient
for female labor-force participation falls by half in magnitude, it remains statistically significant at a 1-percent level.26

The reform and opening policy

Another important historical event (probably more important than the OCP) in the late 1970s, the Reform and Opening Policy,
initiated China’s grand marketization reform. China’s GDP per capita grew from $307.1 (in 2010 USD) in 1978 to $7806.9 in 2018,
with an annual growth rate of 8.43 percent (World Bank, 2020). Is it possible that the rapid economic development of the early
1980s caused the decline in the marriage age? For example, a wedding ceremony is a costly event, and increased income allows
marriage at an earlier age. While it is beyond the scope of this study to establish rigorous causality from income to marriage age, we
assert that the income increase is unlikely to explain the decline in the marriage age in the early 1980s. A large body of literature has
documented a robust positive association between economic development and marriage age in China: people marry later in richer or
coastal provinces (Li, 1985; Ji and Yeung, 2014); young people who are either better educated or higher paid get married later (Xu
et al., 2003; Ji and Yeung, 2014). Fig. 1 shows that the trend in the marriage age is the opposite of that of GDP per capita in the
early 1980s. However, the two trends go in the same direction in the late 1980s.27 To summarize, the income increase in the early
1980s is unlikely to explain the simultaneous decline in the marriage age.

4.4. Mechanisms

In this subsection, we provide some suggestive evidence regarding how later marriage reduces men’s number of children and
encourages female labor-force participation. We are also interested in understanding the gender-heterogeneous effects—e.g., why
later marriage reduces men’s number of children but not women’s. Our discussion here is suggestive and incomplete because of data
limitations. An RPJK design requires a large data sample because it only uses the observations within a short window around the
cutoff cohort. Census data is the only candidate for accurate estimation, to the best of our knowledge. However, one well-known
limitation of censuses is that they contain relatively few variables.

Men’s number of children

It is intuitive that later marriage reduces births: the best ages for fertility are limited and do not respond to policy changes. Later
marriage mechanically shrinks the optimal period for childbirth because out-of-wedlock births were rare in China at the time (Xie,
2013). A more challenging question is why later marriage reduces men’s number of children but not women’s. One likely answer is
that later marriage also changes the matching process of marriage. Panels (a) and (b) in Fig. 7 show that later marriage reduces the
age gap (the husband’s age minus the wife’s age) by approximately 1 to 1.5 years. One’s ability to have children not only depends
on one’s age but also on the age of one’s spouse. This fact implies a double penalty of later marriage on men’s ability to have
children: as men get older, their wives get older by an even greater amount. On the contrary, from the women’s perspective, their
husbands are relatively younger when they get married late, which partially explains why the negative effect on births is weaker
for the female sample.
Another potential reason why later marriage in men reduces the number of children whereas later marriage in women does not
is the balance of bargaining power in the marriage market.28 If later marriage improves women’s bargaining position and a wife
wishes to have fewer children than her husband (Ashraf et al., 2014), we will observe a decrease in fertility from men’s perspective
and no change from women’s perspective. We find some evidence that the matching pattern in the marriage market does change
in response to an unexpected relaxation of late-marriage requirements. Even if women are not entirely free to choose the timing of
their marriage because traditional Chinese culture encourages young people to get married early (Thornton and Lin, 1994),29 they
(or their parents) can still choose whom to marry. It is fair to assume that women prefer younger and better-educated husbands.
Using the RPJK design, Fig. 7 shows that the husbands of later-married women are approximately 1–1.5 years younger and receive
0.5–1 more year of education.
Additionally, if the bargaining story partially accounts for the differential results in the number of children for men and women,
we expect the difference to be larger in places where the male-to-female sex ratio is more unbalanced (Weiss, 1997). Because a

25 Gong et al. (2015) identify cohorts born between September 1946 and August 1960 as the most affected cohorts according to the nature of the send-down

movement.
26 Appendix C partly explains this gap. The kink in female labor-force participation can be decomposed into a kink in the performance of household duties

and a kink in early retirement. The concept of ‘‘retirement’’ does not apply to most rural people in China because one has to be employed in the formal sector
to be eligible for retirement.
27 Note that our RPJK design estimates some local effects driven by the policy shift in late-marriage regulations. Therefore, what happened in the late 1980s is

not relevant to our empirical strategy. Existing studies attributed the steady increase in the marriage age since the 1980s to China’s rapid economic development
during the period (Guo, 2003; Ji and Yeung, 2014; Yu and Xie, 2015; Xu, 2019).
28 We thank one referee for suggesting this point.
29 Economic incentives also encourage youths (especially rural youths) to marry early. Two well-recognized economic benefits of marriage are the division of

labor to exploit comparative advantage and risk pooling (Weiss, 1997). Those benefits are particularly valuable in an agricultural setting where farmers have
no insurance and face considerable income uncertainty.

728
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Fig. 7. Effect of later marriage on the age gap and education gap between husband and wife. Note: This figure shows the effect of later marriage on the
husband–wife age/education gap from men’s and women’s perspectives, respectively. The graphs present the results of the second-stage estimations (along with
the 95 percent confidence interval) using local linear regressions with alternative bandwidths in one-month increments from 12 to 48 months. The vertical dash
line refers to the CCT optimal data-driven bandwidth.

vast majority of marriages in China take place locally, we calculate the prefecture-level local sex ratio for the marriage market as
the ratio of the number of men aged 23–26 years to that of women aged 21–24 years in a prefecture. We calculate the local sex
ratio using the 1990 population census, and Appendix Table A6 reports the results for high and low sex-ratio areas, respectively.
The results are consistent with a bargaining story. The gender difference in results for the number of children is particularly large
(−0.593 for men versus 0.206 for women) in high sex-ratio areas where women are supposed to have more bargaining power. The
gender difference is significantly lower in low sex-ratio areas (−0.421 versus −0.143).

Women’s labor-force participation

The gender heterogeneity effect on labor-force participation is easy to explain: men’s labor-force participation is almost universal
(over 97 percent, Table 1). Therefore, the key is to understand why later marriage encourages women to participate in the labor
market. One immediate consequence of later marriage is later childbirth. Appendix Figure A3 uses the 1990 census and suggests
that later marriage does translate into later childbirth.30
Motherhood delay is usually believed to be beneficial to women’s labor-market performance. For example, Miller (2011) find
that the delay of motherhood leads to a substantial increase in earnings of 9 percent per year of delay (6 percent resulting from the
increase in working hours and 3 percent from the increase in the hourly wage rate). Troske and Voicu (2013) note that delayed first
birth reduces the negative effect of the first child on labor supply. Using miscarriages and stillbirths as sources of identification, Bratti
and Cavalli (2014) also find that delayed first birth raises the likelihood of a woman participating in the labor market.

30 Because the census only surveys children in the household, we can only compute the variable (the woman’s age at first childbirth) if her first child lives

with her. This assumption is more likely to hold in the 1990 census than in the 2000 census. For this reason, we use the 1990 census instead of the 2000
census in Appendix Figure A3.

729
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

We also try to understand the source of greater female labor-force participation. The reasons for not participating in the labor
market include attending an educational institute (or awaiting school enrollment), performing household duties, retirement, and
disability. Appendix C shows that women who married later were less likely to be housewives in both 1990 and 2000 and were less
likely to retire in 2000 when they approached the retirement age. Those observations are in line with the idea that later childbirth
(along with later career interruption) benefits women’s labor-market performance. We discuss more details in Appendix C.

5. Conclusion

Despite the prevalence of the rising age at first marriage all over the world, there is limited evidence regarding the causal effects
of marriage delay in adulthood. We estimate the effects of marriage delay on subsequent outcomes (including fertility, education, and
labor-force participation) by exploiting a missing piece of the history of China’s family planning policies that was largely overlooked
by the previous literature. Before the well-known OCP, China had already started to implement stringent family planning policies in
the early 1970s. Although the fertility restrictions of the LLF policy were not as rigid as those of the OCP, they additionally imposed
requirements on marriage—the ‘‘later marriage’’ policy recommended a minimum marriage age of 25 years for men and 23 years
for women. The late-marriage requirements were effectively relaxed when China’s family planning policies switched from the LLF
policy to the OCP. With the implementation of the OCP, the government put greater emphasis on directly controlling fertility and
loosened the restrictions on marriage. The average age at first marriage rose steadily throughout the 1970s but turned around in
the early 1980s. This policy shift provides us with an opportunity to apply an RPJK design to estimate the effect of later marriage.
Using data from the 2000 census, we find that later-married men have fewer children and that later-married women are more likely
to participate in the labor market. We also find some suggestive evidence that later marriage improves women’s bargaining power
in the marriage market. For example, women were matched with younger and better-educated husbands.
While we try to establish a causal interpretation of later marriage in this study, we have to emphasize that our RPJK strategy,
which exploits the policy transition from LLF to OCP, estimates some ‘‘local’’ effects. Readers need to be cautious about the external
validity of our findings. Two aspects reflect the locality of the causal effects. First, China in the 1970s and 1980s was a poor
agricultural country in East Asia, and the rural population generally held a strong preference for sons (Li and Zheng, 2009).
Additionally, at the time, China was experiencing rapid transitions in terms of both economics (from a planned economy to a market
economy) and demographics (the total fertility rate declined from 5.7 births per woman in 1969 to 2.5 in 1989). Consequently, our
results are probably more applicable to developing countries with cultures similar to China’s. Second, because the policy specifically
designated 25 (23) years as the cutoff age of late marriage for men (women), what we estimated is essentially the effect of getting
married after 25 (23) instead of the effect of marriage age in general.
Our study has its limitations. We are certainly interested in outcomes beyond the three variables (number of children, education,
and labor-force participation) and would like to explore more potential mechanisms. Unluckily, a regression-discontinuity-style
design requires numerous observations around the cutoff to obtain an estimate with meaningful precision. Because of this
requirement, censuses are probably the only viable sources of data for our RPJK design, and our choice of outcome variables and
ability to deeply explore the mechanisms are greatly restricted. We hope that new data can fill this void in the future.

Appendix A. Supplementary data

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.jce.2022.04.005.

References

Ando, M., 2017. How much should we trust regression-kink-design estimates? Empir. Econ. 53 (3), 1287–1322.
Angrist, J.D., Krueger, A.B., 1992. The effect of age at school entry on educational attainment: an application of instrumental variables with moments from two
samples. J. Amer. Statist. Assoc. 87 (418), 328–336.
Asadullah, M.N., Wahhaj, Z., 2019. Early marriage, social networks and the transmission of norms. Economica 86, 801–831.
Ashraf, N., Field, E., Lee, J., 2014. Household bargaining and excess fertility: an experimental study in Zambia. Amer. Econ. Rev. 104 (7), 2210–2237.
Babiarz, K.S., Ma, P., Miller, G., Song, S., 2018. The limits (and Harms) of population policy: Fertility decline and sex selection in China under Mao. working
paper 25130.
Bergstrom, T., Schoeni, R.F., 1996. Income prospects and age-at-marriage. J. Popul. Econ. 9 (2), 115–130.
Blackburn, M.L., Bloom, D.E., Neumark, D., 1993. Fertility timing, wages, and human capital. J. Popul. Econ. 6 (1), 1–30.
Bratti, M., Cavalli, L., 2014. Delayed first birth and new mothers’ labor market outcomes: Evidence from biological fertility shocks. Eur. J. Popul. 30 (1), 35–63.
Bronson, M.A., Mazzocco, M., 2018. Cohort size and the marriage market: Explaining nearly a century of changes in US marriage rates. UCLA CCPR Population
Working Papers.
Buckles, K.S., Hungerman, D.M., 2013. Season of birth and later outcomes: Old questions, new answers. Rev. Econ. Stat. 95 (3), 711–724.
Calonico, S., Cattaneo, M.D., Titiunik, R., 2014. Robust nonparametric confidence intervals for regression-discontinuity designs. Econometrica 82 (6), 2295–2326.
Cameron, L., Erkal, N., Gangadharan, L., Meng, X., 2013. Little emperors: Behavioral impacts of China’s one-child policy. Science 339 (6122), 953–957.
http://dx.doi.org/10.1126/science.1230221, arXiv:http://science.sciencemag.org/content/339/6122/953.full.pdf.
Card, D., Johnston, A., Leung, P., Mas, A., Pei, Z., 2015a. The effect of unemployment benefits on the duration of unemployment insurance receipt: New evidence
from a regression kink design in Missouri, 2003–2013. Amer. Econ. Rev. 105 (5), 126–130.
Card, D., Lee, D.S., Pei, Z., Weber, A., 2015b. Inference on causal effects in a generalized regression kink design. Econometrica 83 (6), 2453–2483.
Carpena, F., Jensenius, F.R., 2021. Age of marriage and Women’s political engagement: Evidence from India. J. Politics 83 (4).
Chen, Y., Fan, Z., Gu, X., Zhou, L.-A., 2020. Arrival of Young talent: The send-down movement and rural education in China. Amer. Econ. Rev. 110 (11),
3393–3430.
Chen, Y., Fang, H., 2021. The long-term consequences of China’s ‘‘Later, Longer, Fewer’’ campaign in old age. J. Dev. Econ. 151, 102664.

730
Y. Chen and Y. Zhao Journal of Comparative Economics 50 (2022) 713–731

Chen, Y., Huang, Y., 2020. The power of the government: China’s family planning leading group and the fertility decline of the 1970s. Demogr. Res. 42,
985–1038.
Chen, Y., Li, H., Meng, L., 2013. Prenatal sex selection and missing girls in China: Evidence from the diffusion of diagnostic ultrasound. J. Hum. Resour. 48 (1),
36–70.
Cherlin, A., 1980. Postponing marriage: The influence of young women’s work expectations. J. Marriage Fam. 355–365.
Coale, A.J., Feng, W., Riley, N.E., De, L.F., 1991. Recent trends in fertility and nuptiality in China. Science 251 (4992), 389–393.
Coale, A.J., Li, C.S., 1987. Basic Data on Fertility in the Provinces of China, 1940–82. East-West Population Inst.
Dong, Y., 2018. Jump or kink? Regression probability jump and kink design for treatment effect evaluation Unpublished manuscript.
Ebenstein, A., 2010. The ‘‘Missing Girls’’ of China and the unintended consequences of the one child policy. J. Hum. Resour. 45 (1), 87–115.
Fan, J., Gijbels, T.-C., Huang, L.-S., 1996. A study of variable bandwidth selection for local polynomial regression. Statist. Sinica 113–127.
Field, E., Ambrus, A., 2008. Early marriage, age of menarche, and female schooling attainment in Bangladesh. J. Polit. Econ. 116 (5), 881–930.
Ganong, P., Jäger, S., 2018. A permutation test for the regression kink design. J. Amer. Statist. Assoc. 113 (522), 494–504.
Gelman, A., Imbens, G., 2019. Why high-order polynomials should not be used in regression discontinuity designs. J. Bus. Econom. Statist. 37 (3), 447–456.
Gibson-Davis, C.M., 2009. Money, marriage, and children: Testing the financial expectations and family formation theory. J. Marriage Fam. 71 (1), 146–160.
Gong, J., Lu, Y., Xie, H., 2015. Adolescent Environment and Noncognitive Skills. Working paper.
Greenhalgh, S., Li, J., 1993. Engendering Reproductive Practice in Peasant China: The Political Roots of the Rising Sex Ratios At Birth. (57), Population Council.
Guo, W., 2003. Initial analysis on the marriage and childbirth mode in China in the 1990s (in Chinese). Popul. J. (5), 18–21.
Gutiérrez-Domènech, M., 2008. The impact of the labour market on the timing of marriage and births in Spain. J. Popul. Econ. 21 (1), 83–110.
Hicks, J., Hicks, D.L., 2019. Lucky Late Bloomers? Consequences of Delayed Marriage for Women in Rural Western Kenya. SSRN working paper.
Huang, W., Lei, X., Sun, A., 2021. Fertility restrictions and life-cycle outcomes: Evidence from the one-child policy in China. Rev. Econ. Stat. 103 (4), 694–710.
Ikamari, L., 2005. The effect of education on the timing of marriage in Kenya. Demogr. Res. 12, 1–28.
Imbens, G., Kalyanaraman, K., 2012. Optimal bandwidth choice for the regression discontinuity estimator. Rev. Econom. Stud. 79 (3), 933–959.
Ji, Y., Yeung, W.-J.J., 2014. Heterogeneity in contemporary Chinese marriage. J. Family Issues 35 (12), 1662–1682.
Jiang, W., Lu, Y., Xie, H., 2020. Education and mental health: Evidence and mechanisms. J. Econ. Behav. Organ. 180, 407–437.
Lee, D.S., Card, D., 2008. Regression discontinuity inference with specification error. J. Econometrics 142 (2), 655–674.
Lee, D.S., Lemieux, T., 2010. Regression discontinuity designs in economics. J. Econ. Lit. 48 (2), 281–355.
Li, R., 1985. An analysis of the age of first marriage in China (in Chinese). Popul. Res. (1), 28–32.
Li, H., Zheng, H., 2009. Ultrasonography and sex ratios in China. Asian Econ. Policy Rev. 4 (1), 121–137.
Manoli, D., Turner, N., 2018. Cash-on-hand and college enrollment: Evidence from population tax data and the earned income tax credit. Am. Econ. J. Econ.
Policy 10 (2), 242–271.
McCrary, J., 2008. Manipulation of the running variable in the regression discontinuity design: A density test. J. Econometrics 142 (2), 698–714.
Meng, X., Qian, N., Yared, P., 2015. The institutional causes of China’s great famine, 1959–1961. Rev. Econom. Stud. 82 (4), 1568–1611.
Miller, A.R., 2011. The effects of motherhood timing on career path. J. Popul. Econ. 24 (3), 1071–1100.
Minnesota Population Center, 2020. Integrated public use microdata series, international: version 7.3 [dataset]. minneapolis, mn: Ipums 2020.
Oppenheimer, V.K., 1988. A theory of marriage timing. Am. J. Sociol. 94 (3), 563–591.
Peng, P., 1997. Encyclopedia of Chinese Family Planning (in Chinese). China Population Publishing House.
Peter, N., Lundborg, P., Mikkelsen, S., Webbink, D., 2018. The effect of a sibling’s gender on earnings and family formation. Labour Econ. 54, 61–78.
Porter, M., 2016. How do sex ratios in China influence marriage decisions and intra-household resource allocation? Rev. Econ. Household 14 (2), 337–371.
http://dx.doi.org/10.1007/s11150-014-9262-9.
Qin, X., Zhuang, C.C., Yang, R., 2017. Does the one-child policy improve children’s human capital in urban China? A regression discontinuity design. J. Comp.
Econ. 45 (2), 287–303. http://dx.doi.org/10.1016/j.jce.2016.09.001.
Scharping, T., 2003. Birth Control in China 1949–2000. Routledge Curzon, New York.
Thornton, A., Lin, H.-S., 1994. Social Change and the Family in Taiwan. University of Chicago Press.
Tien, H.Y., 1983. Age at marriage in the People’s Republic of China. China Quart. 93, 90–107.
Troske, K.R., Voicu, A., 2013. The effect of the timing and spacing of births on the level of labor market involvement of married women. Empir. Econ. 45 (1),
483–521.
United Nations, Department of Economic and Social Affairs, Statistics Division, 2020. The World’s Women 2020: Trends and Statistics. Technical report.
Wang, C., Wang, L., 2017. Knot yet: minimum marriage age law, marriage delay, and earnings. J. Popul. Econ. 30 (3), 771–804.
Wang, F., Yang, Q., 1996. Age at marriage and the first birth interval: The emerging change in sexual behavior among young couples in China. Popul. Dev.
Rev. 299–320.
Weiss, Y., 1997. The formation and dissolution of families: Why marry? Who marries whom? And what happens upon divorce. In: Handbook of Population and
Family Economics, vol. 1, Elsevier, pp. 81–123. http://dx.doi.org/10.1016/S1574-003X(97)80020-7.
Whyte, M.K., Wang, F., Cai, Y., 2015. Challenging myths about China’s one-child policy. China J. 74, 144–159.
World Bank, 2020. World Development Indicators. The World Bank, Washington, D.C, (producer and distributor).
Xie, Y., 2013. Gender and family in contemporary China. Population Studies Center research report.
Xu, K.Q., 2019. Changing patterns and determinants of first marriage over the history of the people’s Republic of China. Population 74 (3), 205–235.
Xu, L.C., Qiang, C.Z.-W., Wang, L., 2003. The timing of marriage in China. Ann. Econ. Finance 4, 343–357.
Yu, J., Xie, Y., 2015. Changes in the determinants of marriage entry in post-reform urban China. Demography 52 (6), 1869–1892.
Zha, D., 2019. Schooling Expansion and the Female Marriage Age: Evidence from Indonesia. Working paper.
Zhang, J., 2017. The evolution of China’s one-child policy and its effects on family outcomes. J. Econ. Perspect. 31 (1), 141–160.
Zhao, L., Zhou, M., 2018. Do only children have poor vision? Evidence from China’s one-child policy. Health Econ. 27 (7), 1131–1146.

731