Abstract
Using administrative records of home mortgages in Beijing, we show that dual-income households systematically choose to buy homes that are closer to the wife’s workplace. The wife’s commute from the newly purchased home is on average 11 percent shorter by distance than the husband’s. We estimate a discrete home location choice model and find that households derive substantially larger disutility from the wife’s commute than from the husband’s. Through the lens of a simple collective household model, we show evidence that gender commute gap reflects the intrahousehold division of labor and relative bargaining power.
I. Introduction
There are substantial gender differences in various aspects of the labor market. Men and women differ in labor force participation rates and work hours. They work in systematically different occupations and industries. Despite progress over the decades, women still earn substantially lower wages for the same work. There is an extensive body of research that documents and seeks to explain these gender differences (for reviews, see Altonji and Blank 1999; Bertrand 2011).
Among those who work, there is also a substantial gender difference in the length of the commute. In the United States, female workers on average spend 11 percent less time commuting than male workers.1 Among the OECD countries, women’s commutes are on average 33 percent shorter than men’s (Organization for Economic Cooperation and Development 2015, Table LMF2.6). The gap is even larger among those who are married and those with children. Although the gender commute gap is evident in many countries and persistent over time, as has been documented in the urban studies literature (for example, Madden 1981; MacDonald 1999; Crane 2007), it has drawn less attention in the economics literature.
There are two possible explanations for the observed gender gap in commutes. The first explanation relates the gender commute gap to labor supply and job search and proposes that, given the location of residence, the gap is due to women systematically searching and accepting jobs that require shorter commutes. This is the view adopted by most of the few economic studies in the literature (for example, White 1977, 1986; and more recently, Gutierrez 2018; Le Barbanchon, Rathelot, and Roulet 2021; Liu and Su 2020). The second explanation relates to the choice of home location and proposes that conditional on the location of the workplace, women tend to live closer to where they work. Among married couples, this means that households systematically choose home locations that are closer to the wife’s workplace.
Although these two explanations are not mutually exclusive, and both could be driven by women’s relative distaste for commuting, they have different implications for the gender pay gap. If the gender commute gap is driven by women not willing to accept jobs that require long commutes, shorter commutes among women are directly linked to lower earnings because, conditional on where one lives, searching over a larger radius is more likely to result in finding a higher-paying job. A reluctance to commute limits the set of job offers women can choose from. This view is supported by recent studies (for example, Le Barbanchon, Rathelot, and Roulet 2021; Liu and Su 2020), which show that gender differences in the willingness to pay for a shorter commute explains between 10 percent and 20 percent of the observed gender wage gap.
In contrast, if households systematically choose to live closer to the wife’s workplaces, the observed gender commute gap may be a reflection of a household’s collective effort to mitigate the gender differences in the labor market. Commuting is particularly costly for married women. For example, Black, Kolesnikova, and Taylor (2014) argue that because it is often the wife who runs errands for the household (such as grocery shopping and chauffeuring children to school and extracurricular activities), traffic congestion imposes a large toll on married women’s labor force participation. Living closer to where the wife works makes it easier for her to stay in her job and remain in the labor market.
Due to the lack of data, it is difficult to disentangle empirically these two potential explanations. Few data sets contain both commute and labor market information, and those that do are mostly cross-sectional in nature. Cross-sectional data do not allow us to determine whether the observed gender commute gap is due to job search decisions or home location choices.
We investigate whether there is empirical support for the second explanation, that is, whether households systematically choose to buy homes that are closer to the wife’s workplace than to the husband’s. We leverage administrative records of home mortgages in Beijing, which are provided by an anonymous large mortgage lender. When a household applies for a mortgage, detailed information is collected in order to evaluate the borrower’s creditworthiness. A mortgage is usually cosigned by a coborrower, who in most cases is the principal borrower’s spouse. Important for our purpose, the data set includes detailed addresses of the current and new homes and work addresses of both borrowers. From these addresses we calculate the couple’s commute distances from the current and new homes. The data set also includes information on borrowers’ demographic and economic characteristics, such as age, education, and monthly income. Although the data is cross-sectional, the sequence of the events allows us to treat the couple’s jobs as given and test how they choose their new home location given where the husband and the wife work.
Our sample includes heterosexual married couples where both the husband and the wife work from addresses different from their home.2 We select this specific group to focus on the gender gap between men and women whose commutes are defined.
We first document a substantial and robust gender gap in commute distances. The wife’s commute is about 10 percent shorter than the husband’s from both the current and new homes. The gender commute gap is correlated with the gender earnings gap. When the commute gap is accounted for, the pay gap declines by 7 percent.
We then build and estimate a discrete-choice model of home location decisions. We model household’s choice of residential community and take the couple’s places of work as predetermined. We use the random utility model developed by McFadden (1978). A household derives (dis)utility from commutes, as well as price, exogenous characteristics of the community, amenities in the neighborhood, and unobserved community characteristics. In addition, each household’s valuation of choice characteristics is allowed to vary with its own characteristics.
To estimate the model, we adopt a two-step strategy stemmed from the popular estimation procedure proposed by Berry, Levinsohn, and Pakes (1995) and used in previous empirical studies on house choice (for example, Bayer, Ferreira, and McMillan 2007; Barwick et al. 2021). The first step estimates the disutility from commute distances of the husband and the wife, together with the heterogeneous parameters in household preferences and the mean indirect utility of communities that are common to all households. The second step uses the mean indirect utility as the dependent variable and estimates the marginal utility of prices, community characteristics, and neighborhood amenities.
Although we control for a rich set of community characteristics and neighborhood amenities, both commute distances and house price can be correlated with unobserved community and neighborhood features. To address the endogeneity issue, we reconstruct the couple’s commute distances as the average of and the difference in log commute distances. We show that the variable of interest—the difference in log commute distances between the husband and the wife—is essentially uncorrelated with price and virtually any observable community and neighborhood characteristic, so we treat it as exogenous. We construct an instrumental variable for log price following Bayer, Ferreira, and McMillan (2007). The instrument is based on the exogenous characteristics of communities that are located between 2.5 and 5 km from a given community. The underlying assumption is that the characteristics of communities that are some distance away from a home influence its price through housing market equilibrium, but have no direct effect on utility.
Conditional on the average commute distance of a couple, we find that communities that require longer commutes for the wife generate lower indirect utility and thus are less likely to be chosen. To interpret the magnitude of the relative marginal disutility from commutes, consider a simple case where a couple is choosing from otherwise identical communities between their respective workplaces, which we assume are 10 km apart. Choosing a community that is 1 km away from the wife’s workplace and 9 km away from the husband’s workplace, over another that is equidistant to both workplaces, is equivalent to a four log point decrease in price.
We conduct a host of robustness checks, which include incorporating more flexible heterogeneity in preferences by allowing the marginal utility from amenities to vary with household characteristics, introducing random coefficients on price and commute distances, and estimating a conditional logit model. We also convert commute distances into commute time using a household travel survey and an online digital map service. Our conclusion that households place substantially larger disutility on the wife’s commute than on the husband’s holds up to all these checks.
We investigate whether our results are sensitive to job switches after the household takes out a mortgage. If households systematically buy new homes that are closer to the husbands’ future work locations, our results will overstate the gender commute gap from new homes. Using a panel of employer–employee linked records matched to our mortgage sample, we show that there are no discernible gender differences in the rates of job switches or job losses around the time of a new mortgage. The gender commute gap persists up to nine years after a household takes out a mortgage. If anything, the point estimates suggest that after taking out a mortgage, the husband tends to switch to jobs that are farther away from the new home, and the gender commute gap slightly increases.
We build a collective household model to shed light on household location choices and the gender commute gap. The model predicts that the household lives closer to the wife’s workplace when the value of the household’s public good of which the wife has a comparative advantage in production is higher, or when the wife has more bargaining power in the household. To test these predictions empirically, we explore household characteristics that are associated with the value of household production and bargaining power. Consistent with the prediction regarding the household division of labor, we find that the gender commute gap is larger among households that likely have young children. Consistent with the prediction regarding intrahousehold bargaining power, we find that the gender commute gap is higher among households in which the wife’s relative education or income is higher or the wife is listed as the principal borrower on the mortgage.
This paper contributes to a small but growing body of research on the causes and implications of the gender commute gap. Researchers in urban studies have long noticed that women commute less than men and suspected that it is related to the gender wage disparity (for example, Madden 1981; MacDonald 1999; Crane 2007). As better data become available, research interest in this issue has reignited recently. Black, Kolesnikova, and Taylor (2014) find that married women in metropolitan areas with longer average commutes are less likely to be in the labor force. Gutierrez (2018) documents a substantive and persistent gender commute gap in the United States. He goes on to show that the gender wage gap declines substantially once one accounts for the gap in commute length. Using administrative data from France on job search criteria and exploiting a unique institutional feature that requires unemployed workers to report job offers they receive, Le Barbanchon, Rathelot, and Roulet (2021) show that women have a smaller radius in job search and are more willing to accept lower wages for jobs closer to home. Using the American Community Survey, Liu and Su (2020) show that the gender commute gap contributes to between 16 percent and 21 percent of the gender wage gap in the United States. These studies suggest that the gender commute gap and the gender wage gap are connected through the labor supply channel. This paper shows that the household decision on home location also contributes to the observed gender commute gap and should be considered as complementary to existing studies.
This paper relates to a large body of literature on intrahousehold bargaining and gender gaps in labor market outcomes that follows the seminal works by Becker (2009) and Chiappori (1988, 1992). Recent studies find that as more women pursue a career while changes in traditional gender roles are sluggish, working women may be particularly time-constrained between market work and housework (for example, Bertrand, Goldin, and Katz 2010; Cortés and Pan 2019). Goldin (2014) argues that inflexible work arrangements are the “last chapter” to be turned in order to advance women’s careers. Long commutes tighten the time constraint and can be seen as an inflexibility. This paper shows that dual-career households tend to favor an easier commute for the wife. Holding everything else equal, this could potentially improve women’s labor market participation and reduce the gender pay gap.
Our work also connects to an active strand of research on the relationship between homeownership and intrahousehold bargaining in China. Homeownership is regarded as a signal of status and a prerequisite for marriage in China (Wei and Zhang 2011; Wei, Zhang, and Liu 2017). The spouse who is more credited for the home is shown to gain larger bargaining power in the household (Wang 2014), which is consistent with our finding. This paper shows that home-buying is also related to an important aspect of the labor market. It has implications for the gender pay gap in China, which is large and has been increasing since the market reforms, despite substantial progress in women’s education and qualifications (Liu, Meng, and Zhang 2000; Zhang et al. 2008).
Our findings are related to two sets of models commonly used in the labor and urban economics literature. First, in the neighborhood choice models, the home location choices are endogenized. However, these models typically do not consider commuting (for example, Epple and Sieg 1999). In the few exceptions in which commuting is considered, typically only one work location is allowed per household (Bayer, Ferreira, and McMillan 2007; Kuminoff et al. 2012). In contrast, the recent spatial general equilibrium models of cities take commuting seriously (Monte, Redding, and Rossi-Hansberg 2018; Tsivanidis 2018; Severen 2019). However, they typically model individual rather than collective household decisions. The empirical patterns we document suggest that it is worth building a general equilibrium model that incorporates both endogenous home location choices and the collective household labor supply, as is done in the extensive literature on household migration (for example, Gemici 2007).
The rest of the paper is organized as follows. Section II introduces the data and documents basic empirical facts about commuting and residential choices. Section III presents and estimates the discrete house choice model. Section IV builds a simple intrahousehold bargaining model and exploits the potential mechanisms. Section V concludes the paper.
II. Data and Sample
A. Data
Our main data are the administrative records of home mortgages in Beijing issued by an anonymous mortgage lender between 2005 and 2014. The lender is a major player in Beijing’s residential housing market. In 2013, for example, the lender accounted for about 15 percent of all home sales in the municipality.
When a household takes out a mortgage, the borrower is required to provide detailed information on her demographic characteristics, employment status, occupation and industry, monthly salary, and information on other sources of income and assets. The mortgage contract is typically cosigned by a coborrower, whose detailed information is also collected.
Important for our purposes, the data contain detailed addresses of the borrowers’ workplaces, as well as their current and new homes (the one with the mortgage). In the data, 94 percent of men and 84 percent of women are employed and provide a valid employer address. From these addresses, we are able to calculate linear commute distances from current and new homes.
We drop mortgages with a sole borrower and keep only those with a coborrower that is likely the heterosexual spouse.3 To study commutes, we only keep dual-career households in which both the husband and the wife work.4 Our sample includes 125,466 households.
The mortgage data also include detailed characteristics of the new home under mortgage. Each record of mortgage reports the total price and the size of the home. In Beijing and Chinese cities in general, the vast majority of homes are sold as condominiums in residential complexes. A “residential community” (zhuzhai xiaoqu) refers to a collection of residential complexes that are developed and managed together. The mortgage data include community-level characteristics, such as the year it was developed, the floor-to-area ratio, and the total number of units. Homes in the same community share the same amenities and are priced almost uniformly.
We define a community’s neighborhood as an area within a 2.5-km radius. Though the choice of radius is arbitrary, we believe it is an appropriate definition of a neighborhood in Beijing, and the results are robust to different choices of the radius. Characteristics of the neighborhood are collected from various sources. Demographic characteristics in the neighborhood are constructed from the intradecennial population census of 2015. We measure the share of adults in the neighborhood age 25 and older who have a college degree and residents who have a Beijing hukou.5 Locations of public transit stations, as of 2017, are from maps digitized and geocoded by Gu et al. (2021). We construct each community’s distance to the nearest subway station and count the number of bus stops within the neighborhood. We also digitize and geocode a list of key elementary and secondary schools in Beijing. Average year-round air quality of the neighborhood is compiled from data reported by pollution monitoring sites across the city.
We supplement the mortgage data with the 2015 Beijing Household Travel Survey, which covers around 100,000 individuals from 40,000 households, and the 2015 intradecennial population census, which is the first nationwide survey that has a module on commute. Descriptions of these supporting data sets are provided in Online Appendix A.
B. Summary Statistics
Table 1 reports the summary statistics of the main sample. The average age is 35 for husbands and 33 for wives. Households in the sample are better educated and have higher income than the average household in Beijing.6 Around half of the individuals are college graduates, with women only slightly less educated than men. The average monthly income, adjusted to the 2014 yuan, is 6,100 for men and 4,700 for women.
Summary Statistics
Beijing is a sprawling city that has been undergoing large-scale urbanization and suburbanization over the past three decades. The city had a population of about 21 million in 2019, double that of 2000. Workers have long commutes. In our sample, workplaces are on average 17 km away from the city center by linear distance. Men systematically have longer commutes than women. From current homes, men’s average commute is 11.7 km and women’s is 10.4 km. The new homes the households buy are in general farther away from the city center (22 km) than their current homes (20 km). Because jobs are relatively concentrated in the city center, commutes from new homes are correspondingly longer. The average commute distances is 15.3 km for men and 14.1 km for women.
Table 1, Panel C reports the summary characteristics of communities and their surrounding neighborhoods. There are 5,434 unique communities that occur at least once in our sample. Beijing is among the world’s most expensive cities in terms of housing affordability. Even among those who successfully bought a home, the average price per square meter is twice the average household monthly income. To pay off an average-sized home of 85 square meters, it will cost an average sample household more than 30 years even if it dedicates half of its income to pay for the home.
Figure 1 shows the spatial distributions of homes and workplaces among the sample households. Consistent with geographical patterns of a typical city, more and better-paying jobs are concentrated in the city center.7 Jobs for men and women have similar spatial distributions (Panel A). Relatively few households live within 5 km from the city center, while more than half of the new homes are more than 15 km away (Panel B). As a result, commute distances from new homes are higher for both men and women. Conditional on home location, men’s commutes are consistently longer than women’s (Panels C and D).
Spatial Distribution of Homes and Jobs
Notes: Authors’ calculations using the mortgage data. In Panel A, log wage is first residualized by regressing on a set of gender, education attainment, and age bin indicators.
C. Basic Patterns of the Gender Commute Gap
Using the mortgage sample, we describe the basic patterns of the gender commute gap and its correlation with the gender pay gap. The patterns should be interpreted as simple correlations, not causal effects.
In the first four columns of Table 2, we regress log commute distance from the new home on an indicator for male. The results confirm that there is a systematic gender difference in commute distances from new homes. The difference persists as we control for individual demographic characteristics or restrict the comparison to be between the husband and the wife within the same household, husband’s commute is consistently between 10 percent and 12 percent longer than the wife’s.8 Column 5 shows that wives’ commutes from the current home are on average about 10 percent shorter than the husbands’ commutes.
Gender Gap in Commute and Earning
The two remaining columns show the correlation between the gender earnings gap and the gender commute gap, where we regress log monthly salary on the gender dummy and log commute distance. Female workers earn about 30 percent less than male workers with the same qualifications. However, when it is conditional on the log commute distance, the gender pay gap reduces to 27 percent, a 7 percent decline. The magnitude of the decline in the gender pay gap associated with commuting is similar to those found in recent studies.9 In addition, Column 7 shows that the commute distance is positively associated with log salary, which is a common pattern found in many countries. Conditional on the demographic characteristics and district of the home location, a one log point increase in commute distance is associated with a 0.1 log point increase in monthly salary.
Before we proceed to estimating how commute distances affect home choices, we rule out the alternative explanation that the gender commute gap is due to systematic differences in the spatial distribution of jobs by gender; that is, women’s jobs may be systematically distributed closer to residential neighborhoods than men’s jobs.
The first piece of evidence to reject this alternative explanation comes from Panel A of Figure 1, which shows that men’s and women’s jobs are about equally represented across space. If anything, women are slightly more likely to work within 5 km from the city center. Because most households live more than 10 km from the city center, that alone would suggest longer commutes for women.
As a more formal test, we reshuffle communities in our sample and randomly assign them to each household. We then calculate the couple’s commute distances to this “fake” new home and estimate a placebo gender gap using the same specification as in Column 4 of Table 2. We repeat this placebo test 1,000 times. If women’s jobs are systematically located near residential neighborhoods, we would expect women’s commutes from randomly picked new homes to be shorter than men’s. The coefficients associated with the indicator for men are plotted in Figure 2. The placebo gender gaps are bounded within a tight range between −0.008 and 0.003, while the gender gap using the actual home location choices is 0.1. This result suggests that the gender commute gap is due to households’ intentional choices, rather than a mechanical pattern driven by the spatial distribution of jobs and available homes.
Distribution of Gender Commute Gap Using Placebo New Homes
Notes: Density graph of the coefficient associated with the indicator for male in placebo tests using the same specification as in Column 4 of Table 2. For each placebo test, new homes in the sample are reshuffled and randomly assigned to a household. We repeat this placebo test 1000 times. Estimation of the original sample has a coefficient of 0.103, which is well outside the distribution of placebo effects.
III. Estimating Home Location Choices
A. Model of Discrete Residential Location Choice
Buying a home is a complicated decision. Besides commute distances for the husband and wife, a household also considers home features, neighborhood amenities, and price. The distribution of those factors could be correlated with commute distances, potentially making the gender commute gap an artifact of other underlying correlations. In addition, descriptive analyses do not tell us how much households dislike long commutes. We need a framework to quantify these disutilities.
In this section, we estimate a discrete home choice model that follows the residential sorting literature (for example, Bayer, Ferreira, and McMillan 2007; Kuminoff et al. 2012). In the model, a household decides which home to buy after taking into consideration the price, home features, and neighborhood characteristics, some of which are unobserved to the econometrician. Unlike most of the existing literature that does not consider commuting or only allows one place-of-work for each household, we explicitly model the spouses’ commute distances as factors to be considered.
We treat a residential community as the unit of choice. Condominiums within a community share the same community-level features and neighborhood amenities. They are similarly priced. In our sample, community-by-year-of-mortgage indicators absorb 92 percent of the total variation in log price per square meter. Our mortgage data cover a substantial fraction of all home sales in Beijing, so we assume the communities appearing in the data during the year are representative of those that were available in the market. Following the literature, we treat households as price-takers and community characteristics as exogenously given.10 We also assume that communities have ample supply of homes, so the fact that one household buys a home in a community does not exclude other households buying a home in the same community and has no impact on the equilibrium price.
Specifically, each household i chooses community c to maximize its indirect utility 
, where the choice set includes all communities that had at least one transaction in our mortgage data in the three years around the year of observation (the year of mortgage, the year before, and the year after). 
follows the random utility specification as in McFadden (1978):
1
where 
and 
are, respectively, the husband’s and wife’s log commute distance from community c. pc is the log price per square meter.11 There are two sets of observable community-level features. The first set, Xc, includes community and neighborhood characteristics that do not vary across households. Those features include community characteristics, such as green-area ratio and average floor space, and amenities in the surrounding neighborhood, such as access to the public transit and education. The second set, 
, includes community characteristics that vary across households. Commute distances belong to this set, although due to their central role in our analyses, they are spelled out separately. In most specifications, 
contains the community’s distance to the household’s current home as a way to capture the household’s unobserved preference for certain areas of the city.
To capture households’ heterogeneous preferences on choice characteristics, we allow the evaluation of community features to vary with household characteristics, mi, in which we include log household income and husband’s age. 
, for j ∈ {X, p}, is specified as:
2where 
is the sample average of household characteristic m. α0j captures the marginal utility of variable j for an average household.
The error term is divided into a component associated with each community that is valued the same by all households and potentially correlated across communities, ξc, and a household-specific term, ɛci. ξc captures the unobserved quality of the community c. ɛci is an independent and identically distributed shock in preference.
The parameters in the model, αs and βs, are marginal utilities. To evaluate their magnitudes, we follow the neighborhood choice literature and compare them with the mean coefficient associated with log price, α0p. For example, βh/α0p indicates that the effect of a one log point increase in the husband’s commute distance on the household utility is equivalent to a βh/α0p log point increase in price.
It is worth noting that although we write down a “household indirect utility function” as in Equation 1, so far we are agnostic about the decision process inside the household and thus do not take a side with either the unitary or the collective models of household behavior. We expect βw < βh < 0, which could be a result of a singular household decision-maker who favors a shorter commute for the wife, or a result of intrahousehold bargaining between the husband and the wife. In Section IV, we present some empirical evidence consistent with a simple collective household model with division of labor and intrahousehold bargaining.
B. Estimation
Estimation of the model follows the two-step procedure developed by Berry, Levinsohn, and Pakes (1995). We rewrite the indirect utility function as:
3
where δc and 
are, respectively,
4
5
where δc is the mean indirect utility of choosing c, which is the part of the utility provided by community c that is the same to all households. 
captures the part of the utility that is household-specific. θδ consists of preference parameters in the mean indirect utility and θλ consists of heterogeneous preference parameters in 
.
Assuming that 
is drawn from the extreme value distribution, the probability of household i choosing community c is
6
The first step of the estimation uses a maximum likelihood estimator (MLE), which estimates the heterogeneous parameters θλ and the vector of mean indirect utility δ. The MLE maximizes the probability that the model correctly matches each household with its chosen community. The log likelihood function is
7
where 
is an indicator that is equal to one if household i chooses community c in the data and zero otherwise.
To estimate δc, we follow the nested contraction mapping procedure as in Berry, Levinsohn, and Pakes (1995). To further reduce the computational burden, we randomly select a subsample of 20 available communities (19 randomly sampled communities not selected by the household plus the chosen community) to construct the choice set.12 Online Appendix B provides more details of the estimation procedure.
With δc estimated in the first step, the second step of the estimation decomposes δc into observable and unobservable components according to Equation 4, which we estimate using a linear regression.
C. Addressing Empirical Challenges
There are several challenges to the identification of key model parameters. First, despite that we control for a wide set of community and neighborhood characteristics, housing price could still be correlated with unobserved amenities; that is, E[ξcpc] ≠ 0. To the extent that households value amenities and that amenities are capitalized in the house price, the marginal utility of price will be overestimated (or the marginal disutility of price underestimated). Estimating the marginal (dis-)utility of price is important because it will be used to gauge the magnitude of disutilities from commute distances.
Following Bayer, Ferreira, and McMillan (2007), we construct an instrumental variable for the price of an community using the characteristics of surrounding communities that are close enough to affect the price of the community (relevance), but far enough such that their characteristics are uncorrelated with the amenities that residents of the community enjoy (validity). We assume the “neighborhood”—whose amenities are close enough for the residents of the community to place some value on—as an area that is within a 2.5-km radius of the community, and the “local housing market”—within which the relative demand and supply factors affect the price of the community—as an area within 5 km. Therefore, average exogenous characteristics of the surrounding communities between 2.5 and 5 km away from the community of interest serve as instrumental variables.13 We construct a stronger instrument by aggregating these exogenous characteristics into a predicted price using a hedonic price model. Specifically, we first regress log price on exogenous characteristics for all communities in our sample, and we use the estimated coefficients to predict a price based on the average exogenous characteristics of communities inside the local housing market but outside the neighborhood. The predicted log price serves as the instrument.
The second challenge to identification has to do with the endogeneity of commute distances. It is unclear how commute distance is correlated with the unobserved utility of the community. On the one hand, both amenities and jobs are overrepresented in the city center, and commute distance can be negatively correlated with amenities. On the other hand, communities that require longer commutes are more likely to be found in the suburbs; they tend to be newer, more spacious, better designed, and have better facilities.
We circumvent the potential endogeneity problem by reconstructing commute distance variables as the average of and difference in log commute distances:
8
9
Online Appendix Table E.2 shows that 
is essentially uncorrelated with 
, as well as any community and neighborhood characteristic, including price. This is intuitive, as, conditional on the average commute distance, the allocation of commute distances between the husband and the wife is likely an idiosyncratic factor for the household and should not be systematically correlated with community and neighborhood characteristics. Therefore, we treat the gender difference in commutes as exogenous.
There is some evidence that average log commute is positively correlated with appealing community features but negatively correlated with neighborhood amenities. The coefficient associated with the average log commute, which may be biased, is not the key parameter of interest. Nevertheless, as we briefly discuss in Online Appendix C.3, the bias is likely not severe and does not affect the estimation of log price.
With the construction of commute distance variables, Equation 1 becomes
10
We are primarily interested in βd/β0p.
D. Baseline Results
Table 3 reports results from estimating various versions of the first step. Column 1 is the most parsimonious specification, in which we only control for 
and 
. Households derive negative utility from long commutes. Conditional on the average commute distance, a longer commute for the husband relative to the wife is associated with a higher utility. In the remaining columns, we include log distance between the new home and the current home, which captures unobserved household preference for certain areas of the city, and gradually add interactive terms between community and neighborhood characteristics and household characteristics to account for heterogeneous preferences. Once we control for distance to the current home, the coefficients associated with the two commute distance measures remain essentially unchanged with the addition of interactive terms.
House Choice Model—First-Stage Estimates
With the average community utility δc estimated from the first step, the second step uses it as the dependent variables and runs a linear regression on community and neighborhood characteristics. Table 4 reports the results of those estimations. Two-stage least-squares (2SLS) regressions that address the endogeneity of housing price lead to larger estimates of the disutility of price than the ordinary least squares (OLS) regressions, which is consistent with the direction of the bias we would have expected. Results are remarkably similar with δc values from different first-step models.
House Choice Model—Second-Stage Estimates
Some of the coefficients associated with various amenities flip signs from the OLS to 2SLS estimations. Their signs and magnitudes in the 2SLS regressions are mostly intuitive. However, it is important to note that the amenities are not the key variables of interest, and their identification may still suffer from various issues, including multicolinearity, measurement errors, and omitted variable biases. For example, the positive sign associated with management fee probably indicates that it captures the effect of some unobserved community quality. The inclusion of the amenities helps gauge the validity of the instrumental variable. In fact, the 2SLS estimates of the coefficient associated with log price is robust to whether or how the amenities are included (see Online Appendix Table C.2 for details).
Since results are similar across all models, we use Model 2 for the remainder of the analyses. In this specification, βd/β0p is 0.016, which suggests that holding the couple’s average commute distance constant, a one log point increase in the husband’s commute distance relative to the wife’s is equivalent to an increase in price by 0.016 log points. To give a more concrete example, suppose the couple’s workplaces are 10 km apart, and the household can choose to live in any point on the linear line between the two workplaces. Choosing a home location that requires 1 km of commute for the husband and 9 km for the wife over one that has an equal commute distance for the husband and wife is equivalent to an increase in price by four log points.
E. Robustness Checks
This section presents a few robustness checks to the baseline results. Additional checks are presented in Online Appendix C.
1. Commute time
Thus far, we have measured the length of commute as the linear distance between work and home. However, commuting is costly not only because workers need to travel many miles but also, and arguably more importantly, because it takes time. Due to the layout of the road network and the endogenous adoption of different modes of transportation, commute distance may not map proportionally to commute time. We impute commute time and modes of transportation using the 2015 Beijing Household Travel Survey. We briefly describe the imputation below but relegate the details to Online Appendix C.1.
The Travel Survey covers around 100,000 individuals from 40,000 households in Beijing. The Survey includes a travel diary that tracks each respondent’s whereabouts on the day of survey. For each trip, the diary records the departure and arrival locations and time, the purpose of the trip (from which we define commutes), and the modes of transportation used. From the travel diaries, we compile 62,697 commuting trips from 28,366 workers. Commute time can be directly calculated as the difference between the end time and the start time of the trip. For commute distance, the survey does not record the route distance or detailed home or work addresses. The smallest identifiable geographic unit is a Transportation Analysis Zone (TAZ), which has a median size of about one square kilometer. The commute distance is calculated as the linear distance between the home TAZ and the work TAZ. We classify commutes into two categories by modes of transportation—cars and public transit, which includes bus and subway.
For commutes in each mode k, k ∈ {car, transit}, we run the following regression relating commute distance and time:
11
where lnCommTimeik and lnCommDistik are, respectively, log time and distance of commute i that uses mode k. 
is a set of fixed effects indicating home and work district pairs (totaling 324 fixed effects). These fixed effects capture heterogeneity in road infrastructure and congestion levels across different parts of the city—for example, traffic speed should be higher between two suburban districts than between two downtown districts. τdow,h is the fixed effect indicating a specific hour on a specific day of the week when the trip started. It captures the varying traffic conditions that may affect the relation between distance and travel time.
We estimate ρ1k to be around 0.5 for both modes of transportation, which suggests that commute time grows only half as rapidly as commute distance. This is due to two reasons. First, all commutes require a fixed cost, such as walking to the transit station or access to one’s car, which is diluted as one travels longer distances. Second, longer commutes typically include a larger portion in the suburban area, where traffic speed is likely higher. A similar relationship between commute distance and commute time has been found in the U.S. data (Couture, Duranton, and Turner 2018).
We impute commute time by transportation mode using the coefficients estimated from Equation 11 for all pairs of community and work addresses in the mortgage sample (including commutes from potential communities of choice). For home–work pairs that are less than 1 km apart, we assume that the person bikes to work and impose a speed of 7 km per hour.
Next we predict each person’s choice of transportation mode. Again using the Travel Survey, we estimate a logit model of using automobile for commute based on the individual’s gender, age, education, household income, spousal characteristics, commute distance, and access to public transit. We allow flexible interactions among these characteristics. The estimation results are reported in Online Appendix Table C.1. We then construct the same set of predictors in the mortgage sample and predict each individual’s mode choice. The predicted commute time is calculated as the sum of the multiplication between the predicted probability of choosing either mode and the corresponding commute time.
The gender gap in predicted commute time is about 7 percent, smaller than the size of the gender gap in commute distance. This is driven by two forces. First, the estimation result from Equation 11 shows that commute time increases more slowly than commute distance. Second, our mode-choice model shows that men are more likely than women to use automobiles for commuting, which is in general faster than public transit. Controlling for the mode of transportation, the gender gap in predicted commute time remains at around 5 percent. The average predicted one-way commute time in the mortgage sample is 55 minutes. A 5 percent gender gap translates into men spending about six minutes more in commute per day.
Table 5 estimates the discrete home choice model with commute distances replaced by predicted commute time. The results are similar when we impose either mode of transportation for everyone (Columns 1 and 2) or across different models we use to predict transportation mode (Columns 3–5). A smaller gender gap in commute measured in time, compared with that measured in distance, translates into larger magnitudes of the coefficients associated with the commute gap and the average commute.
House Choice and Commute Time
2. Job dynamics after mortgage
Job changes often accompany home location changes. We measure the gender gap in commute distances between the new homes and the current jobs. It will be a challenge to our findings if households move to new homes in expectation of future job changes. The gender gap in commute distances may be weaker or even reversed if after taking out a mortgage, husbands disproportionately change their jobs to be closer to new homes.
To investigate this alternative explanation, we obtain the monthly administrative panel of employer–employee linked records. The administrative records on employment relationships are derived from monthly compensation records. Each record essentially has three variables: year–month, anonymized employer ID, and anonymized employee ID. The records we have include the monthly employment relationship of all individuals in our mortgage sample, either as a principal borrower or as a coborrower, between 2006 and 2014, provided that the individual was ever employed in the formal sector in Beijing during this period of time.
We match this monthly employer–employee panel to our main mortgage data, which contain more detailed information on the individual and firm characteristics. From the mortgage data we know the detailed addresses of the employers when the household took out a mortgage. We use this information and the employer ID to fill out the work locations in the employer–employee panel. We are able to recover detailed addresses for about 97 percent of the monthly employer–employee pairs in our data.
We first examine how taking out a mortgage affects the probability of job turnovers for men and women. We define a job switch as when a person changes employers from one month to the next, and we define a “job loss” as when a person is in the employer–employee matched panel in one month but not in the subsequent months. Here, job loss is in quotation marks because disappearing from the panel does not necessarily mean that person became jobless. It may indicate that the person did not receive compensation for other reasons, or the person switched to the informal sector. We restrict job losses only to cases in which a worker is not found in any formal employment relationship for at least three consecutive months to rule out any short-term interruptions of employment. The results are not sensitive to the choice of the number of months that define a job loss. We group the monthly observations at the year level and generate two binary variables for whether a particular worker experienced a job switch or a job loss.
Panel A of Table 6 shows that the probability of job switch and that of job loss both decline after obtaining a mortgage. Before obtaining a mortgage, 15.8 percent of men switch jobs and about 9.6 percent lose their jobs in a year, whereas after obtaining the mortgage, the annual probability of job switch and job loss is 11.2 percent and 4.7 percent, respectively. Similarly, for women, the annual probability of job switching drops from 15.2 percent to 10.0 percent and that of job loss decreases from 9.8 percent to 4.5 percent. The decline in job turnover rates after obtaining a mortgage is not necessarily causal. As workers age, the job turnover rate tends to decline (Keane and Wolpin 1997).
Summary Statistics of Employment Dynamics
More formally, we estimate the following equations:
12
13
In Equation 12, yit is a binary variable indicating whether person i experienced a job switch or a job loss in year t. 
is a binary variable indicating the τth year since the household obtained a mortgage (τ can be negative). The coefficient associated with the year before the mortgage was obtained is normalized to zero. λi is the person fixed effect. θt is the year fixed effect. Equation 13 further investigates the gender difference in the employment dynamics. malei is an indicator for men. t ∙ malei captures the linear trend in the employment dynamics that could differ by gender. The main parameters of interest are γτ.
The first two panels of Figure 3 report the results of estimating Equation 12, where the dependent variable is job switch (Panel A) and job loss (Panel B). In each panel, the graph on the left shows the probability of each event as a function of time, relative to the year before the mortgage was taken out. For both men and women, the job switch and job loss probabilities decline, which is natural as workers age. There is no obvious trend break for job switches around the time a mortgage is obtained, while the probability of job loss seems to level off after a mortgage is taken out. This may suggest a “job-locking” effect of mortgage—when serving a mortgage, losing a job is more consequential, and job switches can be more risky. The graphs on the right show the gender differences in each event by year relative to the mortgage (γτ in Equation 13). There are no discernible differences between men and women in job turnover rates both before and after taking out a mortgage. If anything, women’s job loss rates appear to decline relative to men’s after obtaining a mortgage.
Employment Dynamics around Obtaining a Mortgage
Notes: The vertical bars indicate 95 percent confidence intervals.
We then investigate how the gender gap in commute distance evolves after a household takes out a mortgage. Panel B of Table 6 uses a fully balanced panel of individuals for whom we observe six consecutive years of employer addresses since taking out a mortgage. The results show that men’s average commute distances tend to increase slightly after taking out a mortgage. By the fifth year, men’s average commute distance is 2 percent longer than that when the mortgage was taken out. For women, the commute distance is largely unchanged.
More formally, we estimate versions of Equations 12 and 13 but restrict the sample period to the post-mortgage year (the year when the mortgage is obtained is the leave-out year) and replace the dependent variable with the log commute distance between the new home and the year-end work location. Panel C of Figure 3 shows that relative to the year in which a mortgage is taken out, people gradually work in locations that are farther away from their new homes. By the end of the third year after obtaining a mortgage, the commute distances are approximately 1 percent longer, with men’s commute distances increasing more than women’s. The graph on the right shows that compared with the gender commute gap in the year the mortgage is taken out, the gender gap increases by 1 percent by the third year and 1.3 percent by the fifth year. These effects are precisely estimated, but small in magnitude. Overall, the evidence presented in this subsection shows that our main results on the gender commute gap remain robust in the next few years after a mortgage is taken out.
F. Discussions of the Results
Our findings have implications for the gender pay gap. In contrast to explanations that focus on gender differences in labor supply and job search behavior (for example, Le Barbanchon, Rathelot, and Roulet 2021; Liu and Su 2020), which suggest that the gender commute gap is a cause for the gender pay gap, our results suggest that the gender commute gap could come from home location choices. Everything else equal, shorter commutes could potentially improve women’s access to the labor market and thus reduce the gender pay gap.
It is worth emphasizing that our findings do not rule out supply-side explanations. In fact, the gender gap in commutes, either driven by home location choices or job search behavior, could stem from the same fact that women derive a higher disutility from commuting. Women may be more tightly time-constrained between home and work; thus, a long commute is particularly costly for them. Black, Kolesnikova, and Taylor (2014) find that women’s labor force participation decision is more sensitive to the average commute time of the metropolitan statistical area, and the sensitivity is particularly high for women with young children. Our paper does not quantify the relative contribution of either explanation on the observed gender commute gap. Instead, it should be seen as providing a viable alternative explanation to the existing literature that has so far primarily focused on the supply-side explanations.
It is also tempting to try to gauge the contribution of home location choices, which favor the wives’ commutes, to the narrowing of the gender earnings gap. It is not obvious how to identify such an effect in our setting. A household’s choice of living closer to the wife’s workplace is one of many household decisions. The husband may only agree to live closer to the wife’s workplace in exchange for the wife agreeing to take on more housework. In other words, the gender commute gap may reflect the household division of labor and the intrahousehold bargaining. We explore these possibilities in Section IV by looking at differences in the gender commute gap across different types of households.
To evaluate the causal effect of home location on the gender earning gap, we need exogenous variation in women’s commute distances and to know how much a shorter commute increases women’s job retention rates and wages. This question is beyond the scope of this paper, which treats the new home location as an endogenous choice. In an interesting study that is also set in Beijing, Liu et al. (2017) show that wives in households that did not win the lottery to have the permit to buy a car were less likely to be working, while there is no such an effect on husbands.14 In our setting, however, we show in Section III.E.2 that women are not more likely to experience job turnovers than men, even though new homes are on average 40 percent farther from current workplaces.
One concern with our results is that they are derived from a specific population and could lack generality. Indeed, our sample includes dual-income, heterosexual households that bought homes in Beijing and took out a mortgage from a specific lender. Due to the high housing price in Beijing, it is reasonable to imagine that households that have monthly mortgage obligations cannot afford to lose an income. To the extent that women’s labor force participation decisions are more sensitive to commuting, a household may thus choose to live closer to the wife’s workplace.
Nevertheless, the gender commute gap is observed in various subsamples with different demographics (see Online Appendix Table E.4). It is unlikely that the overall pattern is unique to a specific group of households. In Online Appendix A, we show that the gender commute gap is present in other data sources on Beijing, including the 2015 Household Travel Survey and the intradecennial population census, although those data sets do not speak to the cause of the observed gender commute gap.
The systematic gender commute gap is also not unique to Beijing, which is among the largest and richest metropolises in China. The 2015 intradecennial census is the first nationwide survey to include questions on commuting. Using a sample from this census, Online Appendix Figure A.1 documents that the gender gap in commute time is present in essentially all province-level jurisdictions in China. In fact, Beijing has among the smallest gender gap in the country. According to the census data, the estimated gender gap in commute time is about 3 percent in Beijing. At the national level, the gender gap is about 8 percent. These estimates are robust to the inclusion of various sets of fixed effects, including modes of transportation.
IV. Potential Mechanisms
A. Household Model of Home Location Choices
To interpret the empirical findings presented in the previous section, we build an intrahousehold bargaining model to illustrate how a couple jointly determines their home location. Consistent with our empirical setting, the model takes the couple’s work locations as given. We assume that Pareto weights are constant within couples and vary across couples. We use the model to test how couples with different household characteristics, including the value of public goods and the relative bargaining power between husband and wife, value the gender commute gap differently. As Chiappori (1992) and Browning and Chiappori (1998) point out, the collective model assumes that agents are characterized by their own preferences, household decisions are Pareto efficient, and this model can be implemented by assuming that the household has a welfare function that is a weighted sum of the individuals’ private utility functions. Therefore, under the collective model, households choose their home location to maximize the welfare function, and this is consistent with the approach of the discrete choice model.
A household h consists of a husband (m) and a wife (f). The couple jointly decides where to live. We assume that work locations of the husband and wife are fixed, and therefore, the home location determines the commute time for the husband, denoted as tm, and for the wife, denoted as tf. For simplicity, we assume that the couple chooses a home located on the straight line between the two workplaces, and hence, the sum of tm and tf is a fixed number, which we denote as D.15
Each spouse chooses their own private good ci, home production time li, and leisure Li. Individuals derive utility from their own private good (ci), a household public good (Q), and leisure (Li). Production of the household public good requires labor input from the husband (lm) and the wife (lf).
The utility function of the household is
14
μ is the Pareto weight of the wife. αQ captures the preference on public good Q. αi captures individual i’s idiosyncratic preference for leisure Li.
Household public good Q is produced by the husband’s and wife’s time input:
15
Individual time inputs are aggregated to produce household public good via a constant-elasticity-of-substitution (CES) production function. ρ describes the elasticity of substitution between wife’s time and husband’s time. Depending on the specific kind of household public good, either the husband or the wife may have some edge in marginal productivity. γ captures this possible gender difference. We are nevertheless agnostic about the source why γ may be different from one in certain cases.
We assume that each spouse supplies a fixed amount of time at work. This is a reasonable assumption for our sample since virtually all people work full-time. However, the implications of the model do not depend on this assumption. We present an extended version of the model in Online Appendix D.2, where we allow flexible work hours. Here, the time constraint for individual i is:
16
T is the total time net of working hours. T is divided into commute time ti, time spent on producing the household public good li, and leisure Li.
Household budget constraint is:
17
where wt is the labor earnings of spouse i, and b is the monetary cost of commuting.
Conditional on the couple’s work locations, the household’s problem is to choose commute time, home production time, leisure time, and consumption level for the husband and the wife. We provide detailed solutions of the model in Online Appendix D.1.
The model generates the following predictions regarding how the couple’s relative commute time changes with the value o f the household public goods and the couple’s relative bargaining power. Proofs of these predictions are provided in Online Appendix D. Note that the model predicts that a change in the earnings of the husband and wife has no direct effect on commute time. Yet the relative earnings of the husband and wife can affect the couple’s bargaining power, which in turn can affect the spouses’ commute times.
Proposition 1. When the value of public good αQ increases, husband’s commute time tm increases and wife’s commute time tf decreases if 
, while husband’s commute time decreases and wife’s commute time increases if 
.
Because ρ < 1, given preference parameters μ, αf, αm, when γ is sufficiently large, the commute gap between husband and wife increases as αQ increases; when γ is sufficiently small, the commute gap increases as αQ decreases.
Corollary 1. The husband spends more time on household production than the wife, lm > lf, when γ < 1; the husband spends less time than the wife, lm < lf, when γ > 1.
Proposition 2. When the wife’s bargaining power μ increases, the husband’s commute time tm increases; when μ decreases, the husband’s commute time tm decreases.
B. Empirical Evidence
Proposition 1 and Corollary 1 suggest there exists intrahousehold division of labor. As the value of a household public good increases, the spouse who has the comparative advantage in producing that particular public good would spend more time on public good production and have a relatively shorter commute.
The presence of young children generates a large demand for a household product caring for the children, a task that often overwhelmingly falls on the shoulders of the mother. Data from the American Time Use Survey between 2003 and 2019 show that for women between 25 and 54, those with children under age five spend on average 162 minutes per day in “caring for family members,” compared with 36 minutes for those without. In contrast, men with children under age five spend only 91 minutes in the same activity category. According to Corollary 1, this observation is consistent with women having some comparative advantage in this particular task. Proposition 1 thus implies that households with young children would have a larger commute gap between the husband and the wife.
The mortgage data do not include information on family members other than those listed as borrowers, so we do not know whether the household has young children. We focus on households with wives aged 30–39, as this age group is most likely to have young children at home. We reestimate the discrete home choice model with interactions between the difference in log commute distance and whether the wife is between 30 and 39 years old. Column 1 of Table 7 shows that the utility of a larger gender commute gap is larger if the wife is between 30 and 39 years old.16 Admittedly, the finding in Column 1 does not “prove” Proposition 1 because we do not directly observe γ nor spousal time allocation in household production. Many other things may be going on with this age group. Instead, the evidence should be seen as largely consistent with the theory, given the usual observation that demand for household public good is substantially larger with the presence of young children and the evidence from time use data that on average women assume an oversized portion of the task.
Intrahousehold Bargaining and the Gender Commute Gap
Proposition 2 predicts that the gender commute gap is larger in households in which the wife has more bargaining power. We proxy intrahousehold bargaining power by the couple’s relative education, income, and, unique to this setting, who was listed as the main borrower on the mortgage. We divide individuals into two education levels—with or without a college degree—and divide couples into four categories by the couple’s education composition. In a parallel exercise, we divide households into quartiles in terms of the difference between the husband’s and the wife’s income. We expect that the spouse with higher education and income has a larger bargaining power. In addition, we use whether the spouse is a primary borrower of the mortgage as another indicator of the bargaining power. We argue that being the principal borrower is a sign that the spouse played a leading role in buying a home, thus indicating a strong say in important household matters.
We fully interact the gender difference in commute distances with indicators of household categories along the aforementioned dimensions and reestimate the discrete home choice model separately for each set of these interactions. Columns 2–4 of Table 7 report the coefficients associated with these interactive variables.
The first thing worth pointing out is that all coefficients associated with these variables are positive, indicating that all types of households, at least along the dimensions they are categorized, prefer shorter commutes for the wife. On top of this common pattern, households where the wife likely has a larger bargaining power tend to have a stronger distaste for longer wife’s commutes relative to the husband’s. For example, in Column 2, we find that distaste for the wife’s commute is particularly strong in households where the wife is listed as the main borrower of the mortgage. Column 3 shows that, given the husband’s education, households in which the wife is college educated put more disutility on her commute. Column 4 shows that the disutility increases with the wife’s income relative to the husband’s. These findings are consistent with Proposition 2.
V. Conclusions
Leveraging a unique data set of home mortgages in Beijing, we study whether a household’s choice of home location contributes to the gender gap in commutes. We find that households’ new homes are on average 10 percent closer to the wife’s workplace than to the husband’s. We estimate a discrete home choice model and quantify the disutility to the household from the wife’s commute relative to the husband’s. Matching the mortgage data to the administrative employer–employee linked records, we find that women in our sample are not more likely to experience a job turnover than men after buying a new home, despite that moving to a new home on average increases the commuting distance by 40 percent.
Everything else equal, shorter commutes could facilitate women’s access to jobs and improve their labor market outcomes. However, we caution that the home location choice is just one side of the many-faceted division of labor within a household. Our data do not allow us to look directly at intrahousehold time use. We show suggestive evidence that the gender commute gap is associated with the wife’s bargaining power. In addition, our results are restricted to dual-income households in Beijing, although there is evidence that gender commute gap is prevalent across space and various household structures.
Our finding differs from the view that the gender gap in commute is caused by the different job search behavior of men and women. According to this view, women cannot or are unwilling to commute long distances and search only for jobs closer to home, which in turn limits their labor market opportunities.
However, we argue that these two causes of the observed gender commute gap can coexist. To decompose the different channels that contribute to the gender commute gap, future empirical studies should exploit natural experiments in both home location choices and job searches. Due to data limitations, we were also unable to examine the important question of how much the household home location choice mitigates the gender pay gap. Answering this question requires exogenous shocks to workers’ commutes and observing subsequent changes in wages and job retention rates.
Acknowledgments
The authors thank the referees, seminar, and conference participants at Jinan University, Michigan State University, Peking University, Purdue University, and the Midwest International Economic Development Conference. Gu gratefully acknowledges funding from the Peking University Shenzhen Graduate School (Grant No. 1270110213), the Natural Science Foundation of Guangdong Province, China (Grant No. 2020A1515011163), and the National Natural Science Foundation of China (Grant No. 72273008). All remaining errors are those of the authors. The authors do not have any conflicts of interest to disclose. Sources of the proprietary and restricted-access data used in this paper are described in the Online Appendix with instructions for application. Replication programs are available by contacting the authors.
Footnotes
↵1. These numbers are from the authors’ calculations using the 2017 American Community Survey and are based on individuals with ages between 25 and 64 who work at least 20 hours per week. We control for a set of socioeconomic conditions, including a set of education attainment indicators, a set of age group indicators, marital status, whether having children under six years old, a set of metropolitan area indicators, and a set of transportation mode indicators. See Online Appendix Table E.1 for details.
↵2. The data do not explicitly ask the relationship between the borrower and the coborrower; we infer relationship based on the demographic information, such as age and gender. Other household structures identifiable from the data include single-head and multigenerational households. We do not know whether a couple is married or cohabiting, and we are not able to infer whether two borrowers of the same sex constitute a same-sex couple. China does not allow same-sex marriages, but cohabitation of same-sex couples does exist.
↵3. Although the coborrower is customarily the spouse, the data do not indicate the relationship between the main borrower and the coborrower. We focus on pairs of borrowers that consist of a man and a woman of similar ages. We drop mortgages in which the two borrowers are of the same sex or there is an age gap more than 20 years. The focus on heterosexual couples is to exploit gender differences. The restriction on age gap is to rule out mortgages in which the main borrower and the coborrower are of the parent–child relationship. We presume the remaining pairs of borrowers are married heterosexual couples.
↵4. We investigate commutes of workers in other types of households—those with only one spouse working and households of singles—in Online Appendix A.4.
↵5. Hukou is China’s household registration system, which largely ties a person to their place of birth. One may change hukou as one moves to a new place. But obtaining a hukou in biggest cities like Beijing is difficult. Therefore, the share of population with Beijing hukou serves as a measure of the share of nonmigrants.
↵6. See Online Appendix Table A.1 for comparisons of household and individual characteristics across different data sets.
↵7. A large portion of Beijing’s city center is dedicated to museums, parks, monuments, and protected historical neighborhoods. The employment density is highest between 2.5 and 5 km from the city center.
↵8. Online Appendix Table A.3 shows gender commute gaps among the singles, married couples, and those whose spouses do not work.
↵9. Gutierrez (2018) finds that 10 percent of the gender pay gap among childless workers and more than 23 percent of the wage decline attributed to the child penalty can be explained by sex differences in commuting patterns. Le Barbanchon, Rathelot, and Roulet (2021) estimate that around 10 percent of the gender wage gap is accounted for by gender differences in the willingness to pay for a shorter commute. Liu and Su (2020) find that differential commuting choices account for between 16 percent and 21 percent of the gender wage gap.
↵10. Bayer, Ferreira, and McMillan (2007) point out that the exact characterization of the conditions for equilibrium is not necessary for recovering the underlying preference parameters.
↵11. We pool all mortgages in the estimation. Because mortgages in our sample span a period of nine years, we convert housing prices in different years into a constant price by assuming prices of all communities in Beijing increase at the same rate. Specifically, we first regress log unit-area price on community fixed effects and year fixed effects. We then use the estimated year fixed effects to convert all prices to the 2014 price.
↵12. McFadden (1978) shows that in random utility models, estimation using choice-based sampling can asymptotically approximate the true parameters. Results are robust to larger choice sets.
↵13. The exogenous community characteristics include the year of completion, average unit size, green-area ratio, floor-to-area ratio, management fee, and a set of 18 district indicators. The results are not sensitive to a reasonable range of the cutoff radius. Naturally, the instrument tends to be weaker when the neighborhood is larger and the local housing market is smaller.
↵14. In order to reduce traffic congestion, the municipal government of Beijing limits the number of new car licenses issued every month. The new licensees are decided by a lottery. The odds of winning the lottery are low. For example, in October 2018, about three million individuals entered the lottery, and only 6,402 new licenses were issued.
↵15. We assume that residential amenities are the same anywhere in the city. With costly commuting being the only consideration, the household first minimizes the total commuting distance. Any point outside of the straight line between the two workplaces is suboptimal.
↵16. In the Travel Survey, where we have information on children, we find that the gender commute gap for households with children at home is 5.6 percentage points higher than other married households in the sample.
- Received October 2020.
- Accepted September 2021.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.