Content uploaded by Wen-Jen Tsay
Author content
All content in this area was uploaded by Wen-Jen Tsay on Jan 22, 2020
Content may be subject to copyright.
Males’ Housing Wealth and
Their Marriage Market Advantage
C. Y. Cyrus Chu, Jou-Chun Lin, Wen-Jen Tsay
Institute of Economics, Academia Sinica
September 2019
1. Corresponding Author: Wen-Jen Tsay, The Institute of Economics, Academia Sinica, 128 Academia
Road, Sec. 2, Taipei, Taiwan, R.O.C. Tel.: (886-2) 27822791 ext. 296. Fax: (886-2) 27853946. E-mail:
wtsay@econ.sinica.edu.tw
C. Y. Cyrus Chu, The Institute of Economics, Academia Sinica. E-mail: cyruschu@gate.sinica.edu.tw
Jou-Chun Lin, Department of Economics, University of California, Davis. E-mail: joulin@ucdavis.edu
2. We thank Ruoh-Rong Yu and Hao-Chun Cheng for their participation in the early stage of this
project. We also thank Chien-Yu Chen and Chun-Hung Tsang for the research assistance. We are
grateful for the valuable comments of two anonymous referees and the suggestions of the Editor
and Co-Managing Editor of the Journal of Population Economics. An earlier version of this paper
was presented at Oxford University, Paris School of Economics, and Academia Sinica, where useful
comments and suggestions by John Ermisch, Thomas Piketty, and many seminar participants were
greatly appreciated.
1
Abstract
In theory, people who own real estate should have advantage finding a partner
in the marriage market. Empirical analyses along this line, however, face three
issues. First, it is difficult to identify any causality for whether housing facilitates
marriage or expected marriage facilitates a housing purchase. Second, survey
samples usually do not cover very wealthy people, and so the observations are
top coding in the wealth dimension. Third, getting married is a dynamic life-cycle
decision, and rich life-history data are rarely available. This paper uses registry
data from Taiwan to estimate the impact of males’ housing wealth on their first-
marriage duration, taking into account all three issues mentioned above. We find
that a 10% increase in real estate wealth increases probability of a man getting
married in any particular year by 3.92%. Our finding suggests that housing or
real estate is a status good in the marriage market.
JEL classification: C25, J12, R21
Key words: Marriage formation, housing wealth, status good, duration model
2
1 Introduction
Although housing typically is regarded as both an asset and a consumption good, some
scholars emphasize the idea that housing also could be a positional good.1The general
hypothesis is that a man who owns a good housing asset has a much better chance to marry,
or marry earlier, than a man without such an asset. Although this hypothesis is appealing,
the literature has not presented sufficient empirical evidence to support it.
Empirical analyses along this line of thought face three issues. First, it is difficult to
identify any causality for whether housing facilitates marriage or expected marriage facilitates
a housing purchase. Second, survey samples usually do not cover very wealthy people, and so
the observations are top coding in the wealth dimension. Third, getting married is a dynamic
life-cycle decision, and rich life-history data typically are not available to researchers. In this
paper we investigate the interaction between the marriage decision and real estate ownership,
taking into account the three issues mentioned above.
We employ a rare registry micro-panel dataset covering Taiwan to check whether a man
holding more valuable real estate has a better chance to marry earlier than a man without
such an asset, after controlling all socio-economic characteristics. This dataset contains more
than one million observations and covers a life cycle period of 12 years, thus allowing us to
apply the appropriate duration analysis. Furthermore, our data cover the entire population,
overriding the problem of missing observations for the very wealthy.
In order to characterize the dynamic feature of a young man’s marriage decision, we
adopt the methodology of Wrenn et al. (2017) by using a non-linear duration model to
investigate the factors affecting marriage timing. Since assets are particularly important
for a man’s marriage decision, they could correlate with the unobserved preference of an
individual to enter into first marriage. We therefore need to deal with wealth endogeneity in
the duration model. Moreover, we treat parents’ wealth as endogenous by the observation
1See, for instance, Frank (2007), Grier et al. (2016), Marsh (2011), and Wei et al. (2017). Indeed,
spending on status goods is growing in China after the share of males of the total population hit 57% around
2000.
3
that most Taiwanese parents transfer their wealth to their sons, if they could. Therefore, a
young man’s wealth correlates with his parents’ wealth, making parental wealth endogenous
as well.
Although the non-linear duration framework makes instrumentation difficult, we can
adopt the control function approach (Rivers and Vuong, 1988; Papke and Wooldridge, 2008)
proposed by Wrenn et al. (2017) to deal with wealth endogeneity in a non-linear duration
model. In order to identify these endogenous wealth regressors, we utilize two excluded
instrument variables (IV); one is generated by the sudden Taiwan estate and gift tax reform
in 2009, and the other one is the change in the Taiwan stock index level. Both IVs are
exogenous in nature, because they cannot be perfectly predicted in advance. Moreover,
we will show later that they correlate with the magnitude of the wealth regressors and
therefore are legitimate IVs. Our empirical results show that a 10% increase in real estate
wealth increases probability of a man getting married in any particular year by 3.92%. This
marginal effect is over 8 times greater than the effect found if we do not take this endogeneity
into account. Moreover, this housing wealth effect is over 6 times larger than the effect of
financial assets in our robustness checks, suggesting the particular importance of housing
wealth in the marriage decision.
The remainder of this paper is arranged as follows. Section 2 elaborates our research mo-
tivation and links our analyses to the literature concerning the interactions between housing
wealth and marriage. Section 3 introduces the dataset and explains how we construct our
sample and the variables. Section 4 presents the empirical results. Section 5 provides con-
cluding remarks.
2 Background
2.1 Why Study Taiwan and Why Consider Endogeneity
According to a survey of mothers in China with young daughters by Shanghai Daily in March
2010, which Wei et al. (2017) also refer to, 80% of mothers say that they would object to their
4
daughters marrying a man who does not own any form of housing. Taiwan shares a similar
cultural background with that country, and the obsession of owning housing is also strong
among its young couples and parents. Several waves of Taiwan Social Change Surveys2also
indicate that a very high percentage of Taiwanese parents (39.0% in 2015) believe that as
long as they have the capability, they should help buy a house for their children, especially
for their sons. Furthermore, according to the 2015 Panel Study of Family Dynamics (PSFD)
Survey conducted in Taiwan, 48.7% of young couples hope to have an independent home after
marriage, if possible. There is also widespread evidence showing that parents do help their
children buy a house, either by providing the initial downpayments or supporting monthly
mortgage expenses (Chu and Yu 2010, Chapter 10).
The fact that housing ownership is an important determinant of marriage also suggests
a potential consequence of income, wealth, and housing inequality: it is more difficult to
get married if one is relatively poor. In particular, Albouy and Zabek (2016) find that the
distribution of housing values has grown more unequal in recent years than in the mid-1900s.
As housing prices keep surging, which is particularly true in urban areas, the greater unequal
income and wealth distribution make the housing market a battlefield of the rich, thus raising
housing prices even further. Under these circumstances, marriage could be much harder for
those who cannot afford a house, more or less making marriage a privilege for those who can
afford it.
Most studies in the literature regarding the role of home ownership on marriage decisions
employ survey data, deriving a correlation between marriage and wealth, but they do not
consider the possible endogeneity of the reverse causality problem. Specifically, it may be
the case that housing ownership facilitates marriage, or it may also be the case that one’s
intention to get married facilitates the act of buying a house. This paper directly confronts
this endogeneity challenge and looks to test the causal hypothesis of whether the timing of
young men’s first marriage is affected by the market value of real estate they own.
The above-mentioned endogeneity is technically present in the correlation between the
2Source: http://www.ios.sinica.edu.tw/sc/
5
error terms of the marriage and asset acquiring equations. We will explain this in an intuitive
manner. In particular, a man realizes the importance and virtue of accumulating wealth;
however, he also values the freedom of being single. As a consequence, he does not want
to marry very early, but his possession of greater wealth makes him more attractive in
the marriage market than a man without such wealth. In this case, a negative correlation
between his wealth accumulation and his marriage tendency would result in much smaller
effects of wealth on his marriage probability if we do not properly control the endogeneity.
This scenario is helpful for us to analyze the empirical results we will derive in later sections.
In our empirical investigation, aside from young men’s own wealth, we also consider parents’
wealth as an endogenous variable, as it too suffers from reverse causality and unobservable
factor correlations given the dynamics of intra-household transfers. We believe that this
research is the first to use millions of high-quality micro-datapoints to consider parental
wealth endogeneity on marriage decisions.
2.2 Literature Review on Home Ownership and Marriage
To understand what determines people’s marriage decisions and what is delaying them,
previous literature has examined various socio-economic statuses. A rich strand of research
focuses on human capital-related factors such as employment maturity (Oppenheimer et al.
1997; Oppenheimer 2003), stable sources of earnings and income (Burgess et al. 2003; Xie
et al. 2003; Gibson-Davis 2009), and education (Yu and Xie 2015), with researchers finding
correlations with the timing of first marriage. However, perhaps due to data limitations, few
have considered the influence of wealth on marriage, which is certainly a crucial factor of
financial capability that is a concern of unmarried young people.
Dew and Price (2011) include consumer debts, car values, and savings in their prediction
of first marriage, but do not find any significance. Schneider (2011) uses a more complete set
of wealth variables, including home ownership and financial assets, but only notes moderate
predictability from vehicle values and financial assets.
These studies rely on survey data and thus might suffer from an underestimation of
6
wealth values, and they are unlikely to include those on the richer side of national wealth
distribution. Moreover, as mentioned earlier, none of these studies have tried to disentangle
the causal influence of owning assets on the decision for getting married.
Some other studies try different approaches by analyzing macro-data to highlight the
general significance of home ownership on the marriage rate. Bowmaker and Emerson (2015)
employ county-level median housing mortgage per capita income to predict a crude marriage
rate per county and find a negative significance. Their research points to a macro-level
phenomenon of expecting home ownership as a prerequisite of marriage, yet does not rule
out the simultaneity of marriage decisions and housing ownership.
Wei et al. (2017) use regional variations in the ratio of males to females in China to offer
the following statement: if housing is a symbol of economic status, then the more competitive
the marriage market is, the higher are the values of housing that parents will prepare for
their children. Based on collected aggregate regional data, the authors find evidence that a
higher sex ratio does induce people to purchase ever larger and more expensive houses.
Our study also relates to several recent papers using exogenous family wealth shocks on
important demographic decisions. In particular, Laeven and Popov (2017) exploit regional
variations in house price fluctuations in the United States during the early to mid-2000s
to study the impact of the housing boom on young Americans’ choices related to home
ownership, household formation, and fertility. They take American Community Survey
individual-level and household-level extracts from the Integrated Public Use Microdata Se-
ries. However, in both the ordinary least squares (OLS) and the IV specifications in their
Table 9, Laeven and Popov (2017) do not find that the probability of getting married is
higher even for the homeowners residing in a Metropolitan Statistical Area who experienced
a large increase in house prices between 2001 and 2006. As will be shown later, our finding
is opposite to that found in Laeven and Popov (2017). We think the culture difference is
the main reason behind the difference between our results and the studies based on the U.S.
For example, in Section 2.1, we document that there is widespread evidence showing parents
help their children buy a house, either by providing the initial downpayments or support-
7
ing monthly mortgage expenses (Chu and Yu 2010, Chapter 10). Section 2.1 also provides
evidence that treats housing ownership as a precondition of marriage in a Chinese culture.
That explains part of the cultural difference between the Taiwanese covered in this paper
and the Americans studied in Laeven and Popov (2017).
3 Data and Descriptive Statistics
Our observations for empirical analyses are compiled from the Taiwan registration data
provided by the Financial Information Agency (FIA), Ministry of Finance, Taiwan. The
datasets are de-identified and open to all researchers, but are required to be accessed through
FIA’s data center. In this study we use the 2003-2015 encrypted income tax records, wealth
records, as well as family information records including spouse and marriage date. The year
2003 is the earliest data period provided by FIA, and hence we use all the data available
to implement our analysis. Since the objective of this paper concerns the interrelationship
between real estate values and marriage possibility, we need to compile the detailed wealth
accumulation history of individuals. In the following, we explain how to compile such data
in more detail.
3.1 Income, Wealth, and Family Relations Records
The income tax records contain each taxpayer’s comprehensive incomes, including labor
income, dividends, interest, rent, property trading income, etc. The tax is reported by
households, and each member of the household is marked by family status such as the head,
the spouse, and dependent(s). Those who are above 20 years old are required to report their
taxes independently from their parents or other supporters, unless they are registered as
students. Thus, we are able to identify the years of schooling of individuals.
Three sets of wealth records are used in our study: stock shares, houses, and land.
We also retrieve the total amount of deposits through the ‘interest income’ subitem in the
income tax record. We impose a common interest rate calculated from dividing national
8
total savings deposits by total interest income in our data.3The value of stock shares was
originally recorded at face value in FIA; we have adjusted these face values by their yearly
closing prices upon the stocks’ respective ex-right dates, or on July 314if there is no ex-right
date for the companies in question. The prices of unlisted stocks are adjusted by their net
values or remaining face values if their net values are unavailable.
Houses and land are separately recorded and valued by the registration authorities, and
they are on average 40%-60% lower than the true market value. We adjust the land value by
the yearly county-level percentage of undervaluation assessed by the Ministry of Interior.5
Since both house and land are immobile assets, they jointly serve as one’s positional good,
and thus we add them together as real estate and analyze the data based on it.
Family life-history records are updated up to 2015. First marriage dates might be over-
written if a second marriage occurred before 2015. Thus, we resort to the income tax record
and check whether one is the head of the household with a spouse (or the spouse of the
head) to identify a man’s marriage status. According to Income Tax Law §15, married
couples must file their income tax return jointly, and only couples are allowed to file joint
returns. Thus, the first-time joint-return filing year once people turn 18 years old (the legal
minimum age of getting married, Taiwan’s Civil Law §980) should indicate the year they
first married. Finally, we double-check the marriage date from the family record and update
the actual marriage year if the family record shows a different date.
3.2 Data Compiling Rules
Given that the earliest data period we have is 2003 and we lag a total of two periods in dealing
with endogeneity (which will be illustrated in the next section), the oldest cohort that we
can observe is those who were born in 1987, because the legal minimum age of marriage for
males is 18 years old. In Taiwanese society, where males are expected to bear more economic
3Source: https://www.cbc.gov.tw/CPX/Tree/TreeSelect
4We use the last closing price before July 31 if July 31 is on a weekend.
5Source: https://www.land.moi.gov.tw/chhtml/content.asp?cid=14&mcid=194
9
responsibility, we only consider males since they are expected to face more economic pressure
in the marriage market than females. For males who are young, many of them might not
report their income taxes every year, because of their schooling or military service obligation.
We thus collect our observations whenever they appear in income tax records. According
to national statistics published by the Ministry of Interior,6the population of our target
observation sample is 157,981, and we are able to utilize 130,518 (82.62%) of them, along
with the information of their parents’ birth years, income, and wealth.
In dealing with missing income as well as dependent and marriage statuses in the years
when some of these young men do not report income taxes, we impute their income as
zero. We also try other interpolation methods for the missing values, and the results are
qualitatively the same. For the missing dependent status, we fill it in with the next status
information observed in our data. If a person is missing from income tax data for a year,
and we observe their dependent status the next year, then we infer that they were also a
dependent that year.
3.3 Summary Statistics
Table 1 presents the definitions of the variables used in this analysis, and Table 2 presents
summary statistics of these variables.7In Table 3 we list average real estate value, average
financial asset value, as well as parents’ total wealth separately for the two groups of men,
married and unmarried, before they turn 28. This grouping helps us evaluate the relative
importance of various assets on the possibility of entering marriage. In this table, we find
that married people own a higher value of real estate than their unmarried counterparts do,
while unmarried people own higher values of financial assets. However, the effects of various
types of wealth on entering marriage could be confounded by other observable variables and
6Source: https://www.ris.gov.tw/346;jsessionid=B8ACB6E26C89BE93EA69502F5E94C191
7There are some errors in parents’ age in family records. If the difference between father/mother’s age
and children’s age falls outside of [15, 60], then it is considered as a data error and thus dropped from the
observations. Such errors account for about 0.6% of the sample.
10
unobservable interactions among family members. Accordingly, in the next section we adopt
a time-varying duration model to shed more light on such an interrelationship, in order to
identify the relative importance of various wealth on the timing of marriage. This duration
model deals with the endogeneity of various types of wealth based on the control function
approach. We also compare the effects from this duration model with those based on models
that do not control for endogeneity.
4 Housing Wealth and Duration to Marriage
In this section we illustrate an estimation strategy for the duration to exit the single status
from the time a man turns 18, which is the legal minimum age of marriage in Taiwan.
Aside from real estate values, we are also interested in the effect of one’s financial wealth
in order to investigate the different roles of wealth on marriage timing. Both types of
wealth might be subject to endogeneity issues as mentioned previously, and thus we utilize
a duration model with three endogenous regressors: real estate value, financial asset values,
and parents’ wealth.
Following the setting of Wrenn et al. (2017), who consider price endogeneity in a dura-
tion model of residential subdivision development based on the well-known control function
approach, we model the latent utility of individual igetting married at time tas:
U∗
i,t =X⊤
i,t−1α+W⊤
i,t−1β+ui,t,(1)
where Xi,t−1are socio-economic characteristics of individual iat time t−1, including his
income, parents’ income, etc, and Wi,t−1denotes the price of the three endogenous variables:
real estatei,t−1,f inancial assetsi,t−1, and parents′wealthi,t−1. Assets are endogenous,
because they could be correlated with the unobserved preference of an individual to enter
into first marriage. Parents’ wealth being treated as endogenous lies in the observation
that most Taiwanese parents transfer their wealth to their sons, if they could. Therefore, a
young man’s wealth correlates with the size of his parents’ wealth, making parental wealth
endogenous as well.
11
The reason that we lag the explanatory variables Xand Win eq. (1) by one period is
to reflect the facts that our data are yearly based and that it may take some time for people
to realize the asset value of their dating partners.
The three wealth variables are endogenous and are specified as:
real estatei,t−1=X⊤
i,t−2α1+Z⊤
i,t−2δ1+vi,t−1,
financial assetsi,t−1=X⊤
i,t−2α2+Z⊤
i,t−2δ2+wi,t−1,
parents′wealthi,t−1=X⊤
i,t−2α3+Z⊤
i,t−2δ3+ei,t−1,
(2)
where Xi,t−2are the exogeneous variables in the first-stage regressions, and Zi,t−2are the
instrumental variables. We lag both Xand Zone period in this stage for the same reasons
as in the previous one. In the presence of endogeneity, the error terms in eq. (1) and eq. (2)
are correlated and can be written as:
ui,t =θ1vi,t−1+θ2wi,t−1+θ3ei,t−1+ϵi,t.(3)
The non-zero values of the θseries reveal the presence of endogeneity in the corresponding
variables on the probability of entering marriage. For example, the value vin eq. (2) might
represent the magnitude of a young man’s value toward living alone. If a young man who
places a higher value on living alone is less likely to be involved in marriage, then the value
of θ1in eq. (3) is negative, and single-equation models that treat real estate wealth as an
exogenous variable understate its effect on the probability of getting married. On the other
hand, for a man who follows the traditional path of taking on a career and marrying early,
the values of θ1and θ2are positive, and the effects of various wealth on marriage probability
are overstated if the endogeneity is not well controlled.
Using eq. (3), we rewrite eq. (1) as:
U∗
i,t =X⊤
i,t−1α+W⊤
i,t−1β+θ1vi,t−1+θ2wi,t−1+θ3ei,t−1+ϵi,t.(4)
In other words, the inclusions of the error terms, v, w, and e, serve as control variables to
ensure that the errors ϵare uncorrelated with all the regressors in eq. (4). This is the
rationale underpinning the control function approach adopted in Wrenn et al. (2017).
12
Following Wrenn et al. (2017) and assuming joint normality between the errors, we can
model eq. (4) with a discrete time duration model as:
P r(U∗
i,t = 1|Xi,t−1, Wi,t−1, vi,t−1, wi,t−1, ei,t−1)
= Φ [X⊤
i,t−1α+W⊤
i,t−1β+θ1vi,t−1+θ2wi,t−1+θ3ei,t−1+τt−t0
√1−ρ2],
(5)
where τt−t0denotes a set of time fixed effects for forming the baseline hazard. In our anal-
ysis we control age, age2, and age3in each period to serve as the baseline hazard. This
arrangement is legitimate, because we choose the sample from the cohorts all born in 1987,
and this control is equivalent to controlling time fixed effects. We should have mentioned
in the text that personal demographic and income variables very much capture individual
fixed effects. Not mentioning this and directly jumping to a discussion of the fixed effects are
indeed misleading to readers as individual effects are not taken care due to large observation
sizes. We apologize for this negligence. Moreover, the usual first-differencing and within
estimators cannot be used for the non-linear duration model to control for the fixed effects.
The format in eq. (5) reveals that the parameters can be estimated with the usual probit
model, provided that we obtain suitable estimates of the error terms in eq. (2). Due to the
linearity structure in eq. (2), we certainly can estimate these errors using the OLS estimator,
given that we have legitimate instrumental variables Zin eq. (2) to identify the endogenous
regressors. These IVs must not be correlated with the error term in the main equation. To
accomplish this task, we generate one IV from the Taiwan estate and gift tax reform in 2009,
when the Taiwan government suddenly changed the progressive estate tax rates to a single
rate, creating an up to 40% reduction in these taxes. This new tax scheme exogenously
increases the incentive of parents to transfer their properties to their children early, thus
creating an exogenous increase of various wealth for their offspring.
Another source of IV comes from the use of the rate of change in the Taiwan stock
exchange weighted index. As shown previously, an individual’s wealth and his parents’
wealth are endogenous variables, because his intention to marry may be associated with
both. Since stock index changes are related to one’s financial wealth, it fulfills the first
13
requirement as a legitimate IV for a family’s wealth. A change in the stock index evidently
affects one’s investment decisions, but since it is purely exogenous, it does not impact the
marriage decision directly.
We also employ the aforementioned two IVs to interact with his father’s, or his mother’s
salaries, or his parents’ total income as the sources of IVs, because the magnitudes of a
tax shock and a financial shock through stock index changes differ across parents’ income
and salary levels. The justifications are that parents’ income and salaries can help them
accumulate wealth, and thus they correlate with the three endogenous variables and fulfill
the first requirement as IVs. Moreover, these total income and salaries can be treated as
predetermined variables, because they cannot be decided by parents’ discretion; in most cases
they result from the interactions between demand and supply of labor markets, especially
in the short run. Accordingly, the interactions of these two exogenous shocks and the three
predetermined variables not only meet the criteria as IVs, but help us generate more IVs to
identify the three endogenous variables considered in our duration model. Following Wrenn
et al. (2017), we implement overidentification tests to help us determine the optimal IV
model for the first-stage regressions of the proposed control function approach.
4.1 Estimation Results
In Table 4 we first present the estimation results without taking endogeneity into account.
The model in eq. (1) is estimated using a probit model under this circumstance. We find
that the amounts of both real estate and parents’ wealth facilitate marriage, whereas the
amount of financial assets is insignificant. Although it is expected that greater real estate
wealth and parental wealth positively correlate with marriage decisions as the literature has
shown, it is less convincing that financial assets are irrelevant. It also does not make sense
that higher parental income will deter first marriage. We attribute such counter-intuitive
results to a bad model specification that neglects endogeneity issues and will justify such a
conjecture in the next section. Accordingly, we focus on the discussions related to the results
taking care of the endogenous regressors issues.
14
We now consider the case when real estate and financial assets as well as parents’ wealth
are endogenous. The model in eq. (5) is estimated using a probit model. Table 5 presents
the first-stage results for eq. (2), and Table 6 presents the second-stage findings for eq. (5).
Note that the combinations of IVs are guided by the implementation of the overidentification
test discussed above. Since the main equation of interest is estimated with a probit model,
for ease of explanation we present the average marginal effects of various variables. The
standard errors shown in these two tables are based on the bootstrap procedure used in
Wrenn et al. (2017) with 500 replications.8
The first-stage estimation is implemented with the OLS estimator. Most coefficient
estimates meet our expectations and are significant, partly due to the well-chosen model
specification and partly due to the good quality of data with more than 1 million observations.
Among them, we are interested in the testing results about the coefficient estimates of
the IVs. Based on the F-statistics for testing whether the coefficients on the excluded
instruments in the first-stage OLS regressions are equal to zero, we find that our IVs pass
the first-stage exclusion tests significantly. Indeed, all the test statistics exceed 36, suggesting
that the residuals from these first-stage regressions are valid endogenous wealth controls in
the duration model.
To assess the legitimacy of the residuals from the first-stage regressions for the second-
stage duration analysis, we need to perform overidentification tests. Following Wrenn et al.
(2017), in addition to the three OLS residuals and the three corresponding squared residu-
als (accounting for any non-linear impacts, see Papke and Wooldridge (2008)), we add four
of the five excluded IVs to the right-hand side of the duration model and implement chi-
squared tests of the joint-significance of these four excluded variables. The test statistics are
distributed chi-squared with 4 degrees of freedom. Similarly, the overidentification tests are
based on the bootstrap procedure used in Wrenn et al. (2017) with 500 replications to take
into account the residuals from the first-stage regressions as controls in the duration model,
which are generated regressors. The test statistic from the overidentification test is 9.268
8The results in Table 5 and Table 6 are estimated with the Stata code provided by Wrenn et al. (2017).
15
(p= 0.055), which is not significant at the 5% level, supporting the validity of our control
function approach. Moreover, when checking the sensitivity of the excluded instruments on
our analysis in the next subsection, we replicate the previous empirical analysis by adding
an additional IV, i.e., his brother’s real restate. The p-value of the corresponding overiden-
tification test displayed in Table 7 is 0.244. This finding strongly supports the validity of
the excluded instrument variables for our duration analysis.
We now turn to the main results of this paper. Table 6 shows that, after controlling the
endogeneity of the real estate value, a 10% increase in real estate wealth increases probability
of a man getting married in any particular year by 3.92%, which is in sharp contrast to what
Table 4 shows. Financial assets remains insignificant at this stage as we see in Table 4.
Furthermore, the impact from parents’ wealth also increases from 0.005% to 0.0124% for
every 1% increase in its value, even though it changes to be insignificant. The findings are
interesting and reveal the strong influences from one’s own holding of higher values of real
estate. As a consequence, our findings support the assertions of Wei et al. (2017) that
housing or real estate in a broad sense is a status good in the marriage market. Real estate
is a status good in that its marginal effect on the probability of getting married has a much
larger magnitude than that of financial assets, despite the fact that they are both measured
by the same monetary metric.
Before we give an economic reasoning behind the above findings, we should point out that
several coefficients of real estate residual,financial assets residual,parents’ wealth residual,
and their corresponding squared counterparts are significant, supporting the existence of
endogeneity among them. Moreover, the negative sign of real estate residual indicates that
the coefficient of real estate is understated if endogeneity is ignored. On the other hand, the
effects of financial wealth are overstated given the positive sign of financial assets residual
and that the endogeneity of financial wealth is not well controlled. These findings explain
why we have a much smaller positive estimate of the variable real estate in Table 4.
The negative correlation between real estate acquisition and marriage tendency is not
particularly strange once we take the age range of our sample into consideration. The oldest
16
age that we can observe is only 27 (lag one period), which is just about five years after
having graduated from college and no longer than three years after leaving graduate school
if they ever attended one. They must succeed very early, no matter through their own efforts
or not, at their economic well-being in order to acquire real estate at such a young age. In
this sense, it is likely that they endure less pressure on saving before marriage and may even
value marriage less as one of their life goals. After controlling the preferences, the net effect
of real estate still leads to a higher chance of getting married, as the status good hypothesis
suggests.
We now interpret other control variables that we have not yet discussed. As for parents’
age, it appears that older parents are negatively correlated with the marriage hazard, but
also at a decreasing speed. In addition, the positive correlation between income and the
chance of getting married also agrees with the literature (e.g., South, 2001; Schneider, 2011).
Table 6 reveals that the status of being a dependent decreases the marriage hazard signif-
icantly. Note that according to Taiwan’s tax law, individuals aged 20-60 can be designated
as dependents only when they are still in school, mentally or physically disabled, or unable
to support themselves (see Income Tax Law §17). Since the latter two types account for
only 2% of the population,9to a larger extent we capture the dynamic of education status
through the status of being dependents. If a person is still a full-time student, then of course
it is less likely for him to get married. Even if he is not a student, then he must be reliant
on parents’ support to be qualified as a dependent. In either case, the dependency variable
controls the “readiness” of a person to move from being single to being married. Thus, our
findings are intuitively appealing.
We also find that being the only child in the family delays marriage, which is consistent
with the parental psychology of cherishing an only child and parents’ unwillingness to face
an early empty nest. The result is similar to that found in Yu et al. (2012), whose study
was also set against the societal background of Taiwan.
There is also a significant negative effect for being the eldest son on getting married.
9Source: https://dep.mohw.gov.tw/DOS/cp-2976-13825-113.html
17
Eldest sons normally carry more responsibilities on taking care of their parents than their
younger siblings and are thus more likely to live with them (Das Gupta et al., 2003; Lin et
al., 2003). The fact that eldest sons often co-reside with their parents and serve as parents’
closest support makes it harder for parents to let go of them, thus also likely delaying their
first marriage (Jennings et al., 2012).
4.2 Sensitivity Analysis
To check the sensitivity of the excluded instruments on our analysis, in this subsection we
repeat the previous empirical analysis by adding an additional IV, i.e., his brother’s real
estate.10 The rationale is that a man’s brother might hold a similar amount of wealth as he
does, but the brother’s incentive to marry early might not be the same. This test is relevant
in that it can further check the validity of the exogenous instruments chosen in Table 6.
Actually, the results from the Ftests for the significance of these six IVs, and those from the
associated overidentification test support the validity of the control function approach. The
detailed testing results are presented in Table 7. Moreover, the last column of Table 7 lists
the estimation results when taking into account the endogeneity of various wealths, while
the second column displays the findings assuming all these kinds of wealth are exogenous.
Table 7 displays consistent findings with those without using sibling’s wealth in Table
6, except both real estate residual and f inancial assets residual negatively correlate with
the error term in eq. (1). This implies the impacts of real estate and financial assets on the
probability of getting married increase after controlling the endogeneity. However, the real
estate coefficient is over 6 times larger than that of financial assets in our robust checking
studies, suggesting the particular importance of housing wealth in the marriage decision.
10We interact the five instruments in Table 6 with his brother’s real estate.
18
5 Conclusion
Using Taiwan registration data in a time-varying endogenous duration model, this paper
provides the first micro-level evidence on the status good quality of owning housing property
on individuals’ marriage probability. The estimation results also contribute insights into the
impact of owning real estate and financial wealth, as well as other socio-economic variables
including parents’ wealth, on the likelihood of getting married. One of our major results
is that, corresponding to previous literature, owning higher values of real estate induces an
earlier first marriage, regardless of one’s personal preference of entering marriage. Another
novel result is that more financial wealth indicates an earlier first marriage, but with a much
smaller effect.
Although we have rationalized our results by providing econometric analysis for the
effects from wealth without treating endogeneity, we look forward to further investigations
on this subject through other sources of data and methods. A formal model of marriage that
incorporates the relationships of different types of wealth, as well as personal preferences to
marriage outcomes, is also a potential future research topic.
19
REFERENCES
1. Albouy, D. and Zabek, M. (2016). Housing Inequality. NBER Working Paper 21916.
2. Axinn, W. G. (1992). The Influence of Parental Resources on the Timing of the
Transition to Marriage. Social Science Research, 21: 261-285.
3. Bowmaker, S. W. and Emerson, P. M. (2015). Bricks, Mortar, and Wedding bells:
Does the Cost of Housing Affect the Marriage Rate in the US? Eastern Economic
Journal, 41: 411-429.
4. Burgess, S., Propper, C. and Aassve, A. (2003). The Role of Income in Marriage and
Divorce Transitions among Young Americans. Journal of Population Economics, 16:
455-475.
5. Chu, C. C. and Yu, R.-R. (2010). Understanding Chinese Families: A Comparative
Study of Taiwan and Southeast China. Oxford University Press, Oxford, UK.
6. Das Gupta, M., Jiang, Z., Li, B., Xie, Z., Chung, W. and Bae, H. (2010). Why is Son
Preference So Persistent in East and South Asia? A Cross-Country Study of China,
India and the Republic of Korea. The Journal of Development Studies, 40: 153-187.
7. Dew, J. and Price, J. (2011). Beyond Employment and Income: The Association
between Young Adults Finances and Marital Timing. Journal of Family and Economic
Issues, 32: 424-436.
8. Frank, R. H. (2007) Falling Behind: How Rising Inequality Hurts the Middel Class?.
University of California Press, Berkeley, CA.
9. Gibson-Davis, C. M. (2009). Money, Marriage, and Children: Testing the Financial
Expectations and Family Formation Theory. Journal of Marriage and Family, 71:
146-160.
20
10. Grier, K. B., Hicks, D. L. and Yuan, W. (2016). Marriage Market Matching and
Conspicuous Consumption in China. Economic Inquiry, 54: 1251-1262.
11. Jennings, E. A., Axinn, W. G. and Ghimire, D. J. (2012). The Effect of Parents’
Attitudes on Sons’ Marriage Timing. American Sociological Review, 77: 923-945.
12. Laeven, L. and Popov, A. (2017). Waking Up from the American Dream: On the
Experience of Young Americans during the Housing Boom of the 2000s. Journal of
Money, Credit and Banking, 49: 861-895.
13. Lin, I.-F., Goldman, M., Weinstein, Y.-H., Lin, T. and Gorrindo, T. S. (2003). Gen-
der Differences in Adult Children’s Support of Their Parents in Taiwan. Journal of
Marriage and Family, 65: 184-200.
14. Marsh, A. (2011). Social Status and the Demand for Housing. Working Paper. Uni-
versity of Bristol.
15. Oppenheimer, V. K. (2003). Cohabiting And Marriage During Young Men’s Career-
Development Process. Demography, 40: 127-149.
16. Oppenheimer, V. K., Kalmijn, M. and Lim, N. (1997). Men’s Career Development and
Marriage Timing During a Period of Rising Inequality. Demography, 34: 311-330.
17. Papke, L. and Wooldridge, J. M. (2008). Panel data methods for fractional response
variables with an application to test pass rates. Journal of Econometrics 145: 121-133.
18. Piketty, T. and Goldhammer, A. (2014). Capital in the Twenty-First Century. The
Belknap Press of Harvard University Press.
19. Rivers, D. and Vuong, Q. (1988). Limited Information Estimators and Exogenous Test
for Simultaneous Probit Models. Journal of Econometrics, 39: 347-366.
20. Schneider, D. (2011). Wealth and the Marital Divide. American Journal of Sociology,
117: 627-667.
21
21. South, S. J. (2001). The Variable Effects of Family Background on the Timing of First
Marriage: United States, 1969-1993. Social Science Research, 30: 606-626.
22. Wei, S.-J., Zhang, X. and Liu, Y. (2017). Home ownership as status competition:
Some theory and evidence. Journal of Development Economics, 127: 169-186.
23. Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data.
2nd edn. Cambridge, MA: MIT Press.
24. Wrenn, D. H., Klaiber, H. A. and Newburn, D. A. (2017). Confronting Price En-
dogeneity in a Duration Model of Residential Subdivision Development. Journal of
Applied Econometrics, 32: 661-682.
25. Xie, Y., Raymo, J. M., Goyette, K. and Thornton, A. (2003). Economic Potential And
Entry Into Marriage And Cohabitation. Demography, 40: 351-367.
26. Yu, J. and Xie, Y. (2015). Changes in the Determinants of Marriage Entry in Post-
Reform Urban China. Demography, 52: 1869-1892.
27. Yu, W. H., Su, K. H. and Chiu, C. T. (2012). Sibship Characteristics and Transition
to First Marriage in Taiwan: Explaining Gender Asymmetries. Population Research
and Policy Review 31: 609-636.
22
Table 1: Variable Definitions
Variable Description
Marriedit Whether iis married in period t
Ageit Age of iin period t
Real estateit Log of sum of i’s housing and land values in period t. Summed up
before log (thousands).
Financial assetsit Log of sum of i’s stock shares and savings in period t. Summed up
before log (thousands).
Incomeit Log value of i’s comprehensive incomes in period t. Summed up
before log (thousands).
Dependentit Whether iis reported as economically dependent in (his parents’)
income tax record in period t
OnlyiOnly-child dummy
OldestiFirst-born child dummy
Parents’ incomeit Log value of i’s parents’ total comprehensive income in period t.
Summed up before log (thousands).
Father’s salaryit Log value of i’s father’s salary income in period t. Log (thousands).
Mother’s salaryit Log value of i’s mother’s salary income in period t. Log (thousands).
Parents’ wealthit Log value of i’s parents’ total wealth, including house, land, stock
shares, and deposits. Summed up before log (thousands).
D2009tYear 2009 dummy for estate and gift tax reform.
Stock index changetThe rate of change in the Taiwan stock exchange weighted index in
period t.
23
Table 2: Descriptive Statistics - Cohort 1987
Observations = 1,396,919
(130,518 individuals)
15.75% married by 2015
Variable Mean SD Max. Min.
Marriedit 0.0147 0.1204 1 0
Ageit 21.90 3.1354 27 17
Father’s ageit 51.59 5.4131 87 32
Mother’s ageit 48.55 5.0771 87 32
Real estateit 302.36 2532.79 399489 0
Financial assetsit 217.00 3935.86 1709418 0
Incomeit 100.26 315.07 112140 0
Dependentit 0.5567 0.4968 1 0
Onlyi0.0522 0.2225 1 0
Oldesti0.3686 0.4824 1 0
Parents’ incomeit 701.36 3756.75 1433561 0
Father’s salaryit 301.437 619.512 78736 0
Mother’s salaryit 152.866 332.066 46188 0
Parents’ wealthit 12175.75 65101.64 25439975 0
D2009t0.4412 0.4965 1 0
Stock index changet0.1033 0.2943 0.7834 -0.4603
Notes: Financial data are presented in thousands as defined in Table 1, but
without being logged so as to show their magnitude more clearly. Cohort 1987
means the men born in 1987.
24
Table 3: Real Estate and Wealth Values by Marital Status (thousands)
Married Unmarried
Real estateit 309.68 301.23
(2362.46) (2558.16)
Financial assetsit 202.32 219.27
(4313.53) (3874.05)
Parents’ wealthit 12303.38 12155.98
(126102.97) (49305.65)
Observations 187,403 1,209,516
Note: Standard deviations are in parentheses.
25
Table 4: Simple Probit Analysis - Average Marginal Effect
Variables
Fathers’ age -0.00324***
(0.000272)
Fathers’ age squared 0.00245*** a
(0.000238 a)
Mothers’ age -0.00293***
(0.000384)
Mothers’ age squared 0.00215*** a
(0.000369 a)
Income 0.000107***
(0.0000191)
Dependent -0.0132***
(0.000274)
Only -0.00354***
(0.000486)
Oldest -0.00517***
(0.000231)
Parents’ income -0.0000521***
(0.0000198)
Parents’ wealth 0.0000504**
(0.0000245)
Financial assets 0.00000626
(0.0000201)
Real estate 0.000451***
(0.0000191)
Observations 1,396,919
Notes: adenotes the value is scaled up by 100. Standard errors are in parentheses.
*** p < 0.01, ** p < 0.05, and * p < 0.1.
26
Table 5: First-Stage Results of Control Function Approach (OLS Regressions)
Variable financial real parents’
assets estate wealth
Explanatory Variables
Father’s age 0.168*** -0.124*** 0.385***
(0.00984) (0.00845) (0.00858)
Father’s age squared -0.147*** a0.128*** a-0.345*** a
(0.00912 a) (0.00783 a) (0.00795 a)
Mother’s age 0.244*** 0.0148 0.600***
(0.0126) (0.0108) (0.0110)
Mother’s age squared -0.139*** a0.00858 a-0.489*** a
(0.0128 a) (0.0110 a) (0.0112 a)
Income 0.280*** 0.0462*** 0.00202***
(0.000769) (0.000660) (0.000724)
Dependent -0.525*** -0.359*** 0.234***
(0.00966) (0.00830) (0.00843)
Only 0.510*** 0.131*** -0.588***
(0.0177) (0.0152) (0.0154)
Oldest 0.495*** -0.0101 0.305***
(0.00878) (0.00754) (0.00766)
Parents’ income 0.109*** 0.00896*** 0.226***
(0.000830) (0.000713) (0.000724)
Time Fixed Effect Yes Yes Yes
Instrumental Variables
Stock index change 0.101*** 0.0730*** -0.190***
(0.0247) (0.0212) (0.0216)
Stock index change×-0.00970*** -0.00260 0.00375*
Parents’ income (0.00257) (0.00221) (0.00224)
Stock index change×-0.00362 -0.00698*** -0.0295***
Mother’s salary (0.00243) (0.00209) (0.00212)
Stock index change×-0.595*** -0.462*** -0.0276
D2009 (0.0518) (0.0445) (0.0452)
Father’s salary ×0.00157 -0.0395*** 0.0114***
D2009 (0.00102) (0.000877) (0.000891)
F-Statistics 36.8*** 422.5*** 90.6***
Observations 1,396,919 1,396,919 1,396,919
Notes: adenotes the value is scaled up by 100. Age, age2, and age3are controlled in each period to serve
as the baseline hazard and the time fixed effects control in our analysis. The F-statistics are for the joint
hypothesis test of whether the coefficients on the excluded instruments in the first-stage OLS regressions are
equal to zero. Standard errors are in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1.
27
Table 6: Second-Stage Results of the Control Function Approach - Average Marginal Effect
Variable
Father’s age -0.00269*** (0.000292)
Father’s age squared 0.0019*** a(0.00026 a)
Mother’s age -0.00287*** (0.000386)
Mother’s age squared 0.00204*** a(0.000368 a)
Income 2.83e-05 (2.12e-05)
Dependent -0.0121*** (0.000277)
Only -0.00393*** (0.00492)
Oldest -0.00495*** (0.000221)
Parents’ income -8.34e-06 (2.93e-05)
Time Fixed Effect Yes
Parents’ wealth 0.000124 (0.000153)
Parents’ wealth residual 0.000231 (0.000150)
Parents’ wealth residual squared 3.15e-05*** (5.12e-06)
Financial assets -1.98e-05 (0.000118)
Financial assets residual 4.03e-05 (0.000115)
Financial assets residual squared -8.18e-06 (5.49e-06)
Real estate 0.00392*** (0.000514)
Real estate residual -0.00323*** (0.000506)
Real estate residual squared -2.43e-05** (1.05e-05)
Overidentification test 9.268 (p= 0.055)
Observations 1,396,919
Notes: adenotes the value is scaled up by 100. Age, age2, and age3are controlled
in each period to serve as the baseline hazard and the time fixed effects control
in our analysis. Standard errors are in parentheses. *** p < 0.01, ** p < 0.05,
and * p < 0.1. The residuals are estimated from first-stage regressions. The
instrumental variable for stock index change ×D2009 is the excluded instrument
in the overidentificaiton test (4 df) based on Wooldridge (2010).
28
Table 7: Sensitivity Analysis by Brother’s Real Estate (Probit - Average Marginal Effect)
Variables Exogenous Endogenous
Father’s age -0.00339*** (0.000388) -0.00312*** (0.000413)
Father’s age squared 0.00252*** a(0.000345 a) 0.00225*** a(0.00037 a)
Mother’s age -0.00363*** (0.000521) -0.00359*** (0.000529)
Mother’s age squared 0.00281*** a(0.000504 a) 0.00275*** a(0.000507 a)
Income 8.74e-05*** (2.53e-05) 2.25e-05 (2.73e-05)
Dependent -0.0131*** (0.000364) -0.0124*** (0.000356)
Oldest -0.00526*** (0.000302) -0.00530*** (0.000318)
Parents’ income -6.00e-05** (2.63e-05) -3.19e-05 (3.74e-05)
Time Fixed Effect Yes Yes
Parents’ wealth 4.54e-05 (3.25e-05) 2.72e-05 (0.000188)
Parents’ wealth residual 0.000368* (0.000189)
Parents’ wealth residual squared 3.65e-05*** (6.65e-06)
Financial assets 4.78e-05* (2.68e-05) 0.000334** (0.000141)
Financial assets residual -0.000287** (0.000137)
Financial assets residual squared -7.14e-06 (7.00e-06)
Real estate 0.000467*** (2.63e-05) 0.00210*** (0.000572)
Real estate residual -0.00149*** (0.000569)
Real estate residual squared -1.44e-05 (1.43e-05)
Overidentification test 6.694 (p= 0.244)
Observations 823,766 823,766
Notes: adenotes the value is scaled up by 100. Age, age2, and age3are controlled in each period
to serve as the baseline hazard and the time fixed effects control in our analysis. Standard errors
are in parentheses. *** p < 0.01, ** p < 0.05, and * p < 0.1. The residuals are estimated from
first-stage regressions. We interact the five instruments in Table 6 with his brother’s real estate. The
instrumental variable for stock index change ×D2009×brother’s real estate is the excluded instrument
in the overidentificaiton test (5 df) based on the Wooldridge (2010). The F-statistics are for the joint
hypothesis test of whether the coefficients on the excluded instruments in the first-stage OLS regressions
are equal to zero. The F-statistics are 17.6, 292.6, and 40.5, respectively, when the dependent variables
are financial assets, real estate, and parents’ wealth, respectively. The pvalues associated with three
F-statistics are all less than 0.01.
29
A preview of this full-text is provided by Springer Nature.
Content available from Journal of Population Economics
This content is subject to copyright. Terms and conditions apply.
