Housing Prices and Mating Preferences: Evidence from Online Daters Yongheng Deng∗ Yu Qin† Hongjia Zhu‡ This version: May 26, 2016

Abstract This paper investigates the impact of housing prices on mating preferences. Matching the specified preferences of online daters with prefecture city-level housing prices in China, we find that rising housing prices, as a positive wealth shock to local natives, affect mating preferences for the two genders in different ways. Female users have a stronger preference for male home owners, while male users narrow their preferred ranges for females’ height. We also find that housing prices affect the relative preferences of attributes: As housing prices go up, females have stronger preferences on home ownership than on height, while males are the opposite. Keywords: Housing price; mating preference; gender differences JEL classification: R21; J12; O18

∗ E-mail: [email protected], Business School and the School of Design and Environment, Institute of Real Estate Studies, National University of Singapore. † Corresponding author. E-mail: [email protected], Department of Real Estate, National University of Singapore. Qin acknowledges funding support from Ministry of Education in Singapore: R-297-000-129-133. ‡ E-mail: [email protected], Institute for Economic and Social Research, Jinan University, China.

1

They recognized that the changed financial conditions had raised up a social bar between their daughters and the young mechanics. The daughters could now look higher, and must. Yes, must.” — Mark Twain, The $30,000 Bequest

1

Introduction

Wealth is an important determinant affecting dating and marriage decisions. Papers have found a positive assortative matching pattern in the marriage market1 , indicating that rich people tend to date and marry the rich. Among different asset categories of wealth, housing assets usually account for a large share: Owner-occupied housing accounted for approximately one third of total household assets in the United States in 2007 (Wolff, 2010). In China, housing assets account for more than 70% of household assets, according to the nationwide China Family Panel Studies (CFPS) conducted in 2012 (Xie and Jin, 2015). Housing market booms and busts may significantly affect the perceived wealth of home owners and their wealth status relative to non-home owners. In this paper, we aim to understand the impact of housing prices on people’s mating preferences.2 The booming housing market in recent years in a majority of cities in China has provided a positive exogenous wealth shock to the local natives who own (or whose family owns) housing properties.3 We investigate the impact of this positive wealth shock on mating preferences from three perspectives. First, given that people’s mating preferences are assortative, if one perceives himself or herself to be wealthier due to a housing-price increase, is he or she likely to have stronger preferences for wealthier girls or guys? Second, does the impact of housing prices on mating preferences differ by gender? Third, if mating preferences are multidimensional, do housing prices affect the relative importance of different dimensions of preferences? Combining data from one of the largest online dating websites and city-level housingprice data in China allows us to answer these questions. We obtain a comprehensive dataset from one of the largest online dating websites in China, which contains detailed information on the registrants’ attributes and, more importantly, their stated preferences on a partner’s age, height, income, education and home ownership status.4 The recently surging housing 1

See Browning et al. (2014) for a survey. In this paper, people’s mating preferences are their preferences stated on a dating website, which represents their demands regarding a partner’s attributes. Therefore, we use mating preferences and stated preferences interchangeably in the context of the online dating website in the rest of the paper. 3 Based on the calculation from CFPS, the household wealth growth rate from 2010 to 2012 was 18 percent on average. Over half of the wealth growth from 2010 to 2012 was due to the growth of housing assets (Xie and Jin, 2015). 4 For example, one can specify that he or she prefers someone whose age is from 25-30 years, whose height 2

2

prices in a majority of cities in China have provided an exogenous wealth shock to local natives in the cities because either they or their parents are home owners due to the high home ownership rate (more than 90%) in China. Therefore, matching city-level monthly housing prices with the time that the local natives registered on the website allows us to identify the impact of housing prices on their stated mating preferences. We adopt a repeated cross-sectional design. We match the registration date of the local native daters with the housing prices in that month in their hometown from January 2012 to December 2014. We assume that the user profiles reflect their preferences on their registration date because we find that people rarely change their preferences after they register.5 We examine the impact of housing prices (lagged one month) on the mating preferences of the local natives using a Seemingly Unrelated Regression (SUR) approach because the five dimensions of the stated preferences (age, height, income, education and home ownership) are correlated choices of the same individual. In addition to housing prices, we further control for a rich set of users’ attributes and city by year fixed effects. Furthermore, we examine the impact of housing prices on the relative importance of different dimensions of preferences using a multinomial logit model. Our results suggest that local housing prices affect the mating preferences for both genders but in different ways. As housing prices rise, females have stronger preferences for males who earn a higher income and are home owners. Specifically, if housing prices double, the preferred minimum monthly income will grow by approximately 40.5 percent, and the fraction of females preferring home owners will increase by 5.5 percent. In contrast, with increases in the housing price, males will have stronger preferences for females’ height. If housing prices double, the range of height that is specified by male daters will narrow down by 0.38 centimeters. Such effects are more pronounced among local home owners than local natives with other housing arrangements. In addition, by incorporating joint preferences of home ownership and other attributes, we find that the relative odds of specifying stronger preferences for home ownership than for height requirements increase among female daters in a booming housing market and decrease among male daters. To the best of our knowledge, this is the first paper to document the impact of housing prices on people’s mating preferences. Our findings suggest that both genders tend to express stronger preferences for gender-specific desired attributes when they experience positive wealth shocks. Moreover, the unique dataset used in this paper allows us to collect the stated preferences of daters from nearly 300 cities in China; this improves the external validity of the previous literature, which is usually based on small-scale dating experiments is from 160-170 centimeters, and who has a minimum monthly income of 2000 yuan, at least a college degree, and a home. 5 Only 0.2 percent of users changed their profile in a follow-up data collection in three months for the same user cohort.

3

or online dating observations in a very limited number of cities. Lastly, this paper provides policy implications in terms of social mobility. Wealthier females prefer wealthier males, which is further reinforced with housing-price dynamics. As a consequence, housing prices may exacerbate inequality in wealth distribution from the marriage market channel.

2

Conceptual Framework and Related Literature

Wealth plays an important role in the dating process and the matching outcomes in the marriage market. On the one hand, wealth, such as housing assets and cars, has its use value. Kamara (1994) finds that the probability of buying a house decreases for females if they anticipate marrying soon, as they can stay in their future husband’s house. Additionally, the housing burden affects marriage satisfaction and fertility decisions. Nelson et al. (2013) show that the housing burden ratio is negatively associated with marital satisfaction. A recent paper by Dettling and Kearney (2014) finds that a 10,000 USD increase in housing prices leads to 5% increase in fertility rates among home owners due to the equity effect and a 2.4% decrease in fertility rates among non-home owners due to the increase in the housing burden. On the other hand, wealth serves as a status good. Wei and Zhang (2011) find that Chinese parents with a son increase their savings to improve their son’s relative attractiveness in the marriage market, which partially explains the increase in the household savings rate in China via a skewed sex ratio. Wei et al. (2012) further establish the causal link between China’s skewed sex ratio and rises in housing prices in urban China due to the status good feature of housing. In a recent paper, Chang and Zhang (2015) document that local males in Taiwan exposed to high sex ratios due to the inflow of retreated soldiers from mainland China are more likely to become entrepreneurs to make themselves competitive in the marriage market. Fafchamps and Quisumbing (2005) document a similar story in rural Ethiopia: Parents give more wealth to their daughters if doing so improves their matching outcomes in the marriage market. In this paper, we focus on the impact of housing prices on mating preferences. For local natives, rising housing prices serve as a positive exogenous shock to their perceived wealth. How does this wealth shock affect people’s mating preferences on their partner’s wealth (i.e., home ownership), socioeconomic status (such as income and education) and appearance (such as age and height)? The most relevant literature to guide this question is on assortative matching in the marriage market (Browning et al., 2014; Choo and Siow, 2006; Gustaf, 2011; Kalmijn, 1998), i.e., wealthier people tend to marry wealthier people. However, there are two possible explanations leading to such assortative matching patterns. One is that rich people do not have the opportunities to meet and date poor people due 4

to search frictions. The other is that rich people prefer dating rich people. Using online dating data (where search frictions are minimal), Hitsch et al. (2010) suggest that people’s mating preferences are assortative. In a speed dating context, Belot and Francesconi (2013) also find positive assortative preferences along a number of characteristics, with both women and men preferring partners of a similar age, height, and education level. Therefore, if people’s preferences are assortative, increases in wealth may enhance people’s preferences for wealthier partners. Another strand of literature suggests that people’s mating preferences could be gender specific. Hitsch et al. (2010) find that women place about twice as much as weight on income as men do in an online dating context. Fisman et al. (2006) show that in a speed dating context, women put greater weight on intelligence and race and prefer men who grew up in rich neighborhoods, while men responded more to the appearance of women. Ong and Wang (2014) also suggest that women are more interested in high-income men, while men do not exhibit this preference. It is also worth noting that some papers fail to find gender-specific preferences (Belot and Francesconi, 2013; Eastwick and Finkel, 2008). Kurzban and Weeden (2007) find that gender-specific preferences may depend on contexts. To summarize, if people’s mating preferences are gender specific, we may observe different responses for males and females due to wealth shocks. Lastly, wealth shocks may not only change people’s mating preferences along one or many dimensions but also affect the relative importance of different preferred attributes. For example, as one’s housing value increases, will he (she) compromise more on housing preference but demand more on appearance? It could also be the reverse. Chiappori et al. (2012) summarize individual preferences into a one-dimensional index combining various characteristics and study the marginal rates of substitution between different attributes. They find that men may compensate for 1.3 additional units of BMI with a 1 percent increase in wages, whereas women may compensate for two BMI units with 1 year of education. Our online dating data allows us to investigate the impact of wealth shocks on the relative importance of mating preferences as we can observe users’ specified preference along multiple dimensions.

3 3.1

Data and Empirical Method Online Dating Data

We have collected the user profiles from one of the largest online dating websites in China. The dataset contains the user’s profile photo and self-reported attributes, including current residential city, age, height, marital history, education, monthly salary, occupation, home

5

ownership status, hometown, religious beliefs, ethnicity and so on. The website verifies all of these profiles with users’ national identity cards to make sure that they are not fake profiles. The registration time of the users sampled spans from January 2012 to December 2014. Because the website does not charge a monthly subscription fee, inactive profiles can still remain on the website. As shown in Figure 1, the profiles sampled come from 332 cities, covering almost all of the prefecture-level cities in China. We keep in our sample only local natives whose current city of residence is the same as their hometown, for two reasons: First, migrants are self-selected into their current city of residence, a decision that is likely endogenous to the housing prices in their hometown and the city in which they choose to reside. Second, if migrants claim to be home owners, we do not know whether they are home owners in their hometown or in their city of residence because the website does not ask for the location of their properties. Due to these two concerns, we exclude migrants from our analysis to avoid the contamination of the identification of wealth shocks. Table 1 provides the descriptive statistics for the key personal characteristics of online daters in our sample.6 In total, we have 192,450 local natives, including 60,749 females and 131,701 males. Regarding home ownership, 16.5% female locals claim to own a housing property under their own name, while 34.4% male locals claim to be home owners. This is consistent with the social norm in China that male home owners are more competitive in the marriage market as males are assumed to be responsible for preparing a new home for marriage.7 The average ages for male and female daters are 31.47 and 32.15 years, respectively. Approximately 67.8% of the registrants have no history of marriage, while 29.5% have divorced. On average, the daters have 14.14 years of education (12 years for high school graduates and 16 years for college graduates in China) and a monthly income from 3644.3 to 5256.4 yuan (approximately 600-862 USD). A unique feature of this Chinese dating website is that it allows users to specify their partner preferences regarding age, height, education level, monthly salary, home ownership status, and city of residence. The preferences on age and height are expressed as ranges, while the rest of the preferences are specified by choosing one of the multiple categories designed by the website. If users do not want to specify a preference on a certain attribute, they can choose no specific preference. Specifying preferences is costly. It narrows down the pool of candidates a user can date because candidates who do not satisfy a users preferences are not eligible to contact him or her. Table 2 summarizes the online daters’ partner preferences on income, home ownership, education, age and height. In terms of preferences on income, online daters can specify 6

Please refer to Table A1 for a full list of personal characteristics available in the dataset. Another possibility to explain the higher home ownership rate for males is that male daters are more likely to lie in response to this question to make their profiles more competitive. Unfortunately, we have no way to verify their answers. 7

6

the lower bound of the acceptable income of their partners. It is interesting that there are significant gender differences in income preferences. On average, female daters require a minimum acceptable monthly income of 2843.2 yuan, while male online daters require only 675 yuan. Gender differences are even more pronounced in the preference for home ownership; 41.3% of female daters prefer male home owners, while only 6% of male daters prefer female home owners. This pattern is consistent with the finding of Wei et al. (2012) that home ownership increases the competitiveness of males in the marriage market in China. In addition, regarding preferences on education, female users have a minimum requirement of 12.2 years of education, while male users’ requirement is 10.45 years, on average. Regarding preferences on age, females have a higher upper bound (37.89 years) and a higher lower bound (30.32 years) compared to male daters (31.31 years and 24.33 years, respectively). On average, the preferred age has a range of 7.58 years for females and 6.98 years for males. Lastly, the preferred height range is 12.87 centimeters for female daters and 15.59 centimeters for male daters. Figure 2 shows the distribution of specified preferences on housing, income, education, age and height by gender. If we look further into the different genders preferences regarding the reported home ownership status, we find that female home owners have a stronger preference for male home owners (Figure 2(a)). More than half of the female daters (57.8%) who are home owners specify that they prefer male home owners, while 40.5% female nonhome owners prefer home owners 8 . Among male daters, 8% of non-home owners prefer that their female partner own a home, while 5.5% of home owners do. In addition, if we further decompose housing preference by age group, the pattern is the same for both genders: Preferences for home owners increase with age (Figure 2(b)). The fraction of females preferring home owners doubles from ages 18-25 to ages 35-40, while the fraction almost quadruples among male daters from ages 18-25 to ages 35-40. Figure 2(c) shows the preference on income, which is specified as the minimum acceptable monthly income. In general, female online daters have more stringent requirements on males’ monthly income; 82.5% of male daters do not specify preferences on females’ income, while only 39.6% female daters do not specify preferences on males’ income. Furthermore, 26.2% of female daters prefer males who earn at least 2,000 yuan a month (approximately 320 USD); 26.4% of them prefer males who earn at least 5,000 yuan a month (approximately 810 USD) 9 . Figure 2(d) shows the preferences on education by gender. Again, female daters have higher requirements for their partners’ education compared to male daters; 72.6% of males do not have any preferences on females’ education level, while approximately half of females 8 9

Here, we drop the online daters who are not willing to reveal their home ownership status. The average monthly disposable income in China was approximately 2,000 yuan in year 2012.

7

do not have any requirement on males’ education. Of the females, 5.5% prefer males who have at least completed senior high school; 43.5% prefer males who at least have a bachelors degree. In contrast, only approximately 16.2% of male daters prefer females who have a bachelors degree. Figure 2(e) displays the preferences on age, which is measured by the range that the users specify. We use the range here instead of the mean value because the specified age is conditional on the dater’s own age. Therefore, the range that they specify is a better measurement of their preferences because a wider range indicates looser requirements and thus weaker preferences. The results show that there are no stark gender differences in the distribution of the specified age range. The majority of both genders specify an age range between 3 to 12 years. Figure 2(f) displays the preferences on height, which is also measured by range, for the same reason. It seems that females specify a narrower height range than male daters; 41.3% of males specify a height range greater than 20 centimeters, while only 24.2% of females do.

3.2

Housing-price Data

Our housing-price data are obtained from the Soufun Company, a leading real estate data vendor in China (Deng et al., 2011; Deng et al., 2012; Wu et al., 2012). The city-level monthly housing prices are measured by averaging the unit prices of transactions in the primary housing market of each selected city. One salient advantage of this dataset is that it provides the monthly housing-price data for a larger range of cities than other public data sources, such as the National Bureau of Statistics of China. The dataset covers monthly average housing-price (in yuan/squared meters) data for 113 cities from 2012-2014, including cities of different sizes and various locations (as shown in Figure 3). Therefore, it is considered representative of the major real estate market in China. However, the housing-price data we use are not a strictly balanced panel because they are not always available for all 36 months in all 113 cities. It is reported that 1003 out of 4176 city-month housing-price observations are missing from the data. Overall, the average housing price in the selected cities from 2012-2014 was 7473 yuan/sqm. However, it is noteworthy that the housing market in China is spatially heterogeneous. For example, the average housing price in Beijing is 20184 yuan/sqm, which is about seven times that in Xinzhou City. It is worth noting that the housing-price indices provided by Soufun have very high visibility in China. The housing prices from Soufun appear as the top returned result if one searches Beijing housing price or Shanghai housing price in Chinese via main search engines. Therefore, we believe that the housing-price indices by Soufun are the most relevant measure affecting people’s perception of housing prices in China. 8

3.3 3.3.1

Empirical Approach Seemingly Unrelated Regression

Our main empirical analysis adopts a seemingly unrelated regression approach (SUR), where we regress a user’s preferences for different attributes on the lagged housing prices of the user’s hometown city in the month that he or she registered the website. As the same user’s preferences in different dimensions are jointly specified, an SUR approach is more appropriate than independent ordinary least squares equations because it accounts for the correlated error terms across equations. The specification is as follows:

k = α + βk × ln(hpct−1 ) + Yict

X

λjk Xji + ik

(1)

j k where Yict indicates the kth mating preference of individual i, who is from city c and registered on the website in month t. We investigate the impacts of housing prices on five dimensions of mating preferences, i.e., home ownership, income, education, age, and height. We define preference on home ownership as 1 if individual i explicitly states that he/she only prefers home owners and 0 otherwise. Preference on income is defined as the natural logarithm of the lowest acceptable income level. The value is 0 if the online dater states that any income level is acceptable. Preference on education is measured as the lowest acceptable education level, expressed in years. For example, a bachelors degree corresponds to 16 years, while a high school degree corresponds to 12 years. Preference on age is measured as the preferred age range, measured in years. For example, if the dater specifies that he prefers girls aged 18 to 28, then the range is defined as 10 years. A narrower range indicates more stringent preference on age. Similarly, preference on height is measured as the preferred height range, measured in centimeters. Our core independent variable, ln(hpct−1 ), is the one-month lag value of the housing prices of city c at time t − 1. βk represents the average impact of housing prices on the kth preference. We also control a rich set of covariates in vector Xji , including detailed personal characteristics (see Appendix Table A1 for a full list of controlled personal attributes) as well as city-year fixed effects. The error term ik captures the unobserved heterogeneity that influences the stated kth preference of people i. Therefore, we employ the seemingly unrelated regression models to allow the error terms to be correlated across equations.

9

3.3.2

Multinomial Logit Regression

In addition, we are interested in the impact of housing prices on the relative importance of different dimensions of specified preferences. We categorize the preferences for five attributes, i.e., home ownership, income, education level, age and height, into binary categories, i.e., strong and weak preferences. Specifically, we define the preference for home ownership as strong if the user prefers someone who owns a home and weak otherwise. We define the preference for income as strong if the user has a bottom line of 5,000 yuan for monthly income. We define the preference for education as strong if the user prefers at least college graduates. We define the preferences for age and height as strong if the user specifies ranges narrower than the average range among users of the same gender. We conduct robustness checks of the estimation results using different definitions of the outcome variables. We adopt the multinomial logit model, where we use the joint preferences of home ownership and one other attribute as the outcome variable. For example, for the joint preferences of home ownership and income, there are four possible outcomes: The user has strong preferences for both home ownership and income (outcome 1); the user has a strong preference for home ownership but a weak preference for income (outcome 2); the user has a weak preference for home ownership but a strong preference for income (outcome 3); or the user has weak preferences for both home ownership and income (outcome 4). In the multinomial logit model, we are interested in the change in the odds ratio between outcome 2 and outcome 3 due to housing-price changes. The model is specified as follows:

Outcomekict = α + βk × ln(hpct−1 ) +

X

λjk Xji + ik

(2)

j

The dependent variable, defined as Outcomekict , indicates the joint preference of home ownership and another attribute, k, for user i living in city c at time t. In the multinomial logit regression, we select group (0, 1), which implies weak preference on home ownership but strong preference on attribute k, as the baseline group. Therefore, for group (1, 0), the coefficient βk implies that if there is a 1% increase in the housing price, the probability of preferring set (1, 0) relative to (0, 1) would be expected to increase by βk %, holding all other variables constant. In other words, a 1% change in housing prices leads to a βk % change in P (1,0) , suggesting that housing prices change the relative preferences of home ownership and P (0,1) the kth attribute.

10

4 4.1

Results SUR Results

Table 3 presents the main results for Seemingly Unrelated Regressions. The five columns correspond to the five dimensions of specified preferences. Panel A reports the impact of housing prices on the specified preferences of local female daters, while Panel B reports the results for local male daters. All regressions control for more than 20 variables on personal characteristics10 and city by year fixed effect. We use the logged local housing prices in one-month lag as a measurement of housing prices. Due to space constraints, we report only the coefficients on housing prices in the tables. For female local natives, the coefficients on income and housing preferences are significantly positive, indicating that higher housing prices lead female local natives to raise their preferences on their partners minimum acceptable monthly income and home ownership. Specifically, a 1% increase in housing prices leads to a 0.59% growth in the preferred minimum acceptable monthly income and a 0.08% increase in the percentage of female daters who prefer home owners. If we take the housing prices in Beijing as an example, the highest housing price (24384 yuan) is slightly more than two times the lowest housing price (11990 yuan) in our sample period. According to our estimated results, doubling housing prices would lead to approximately 40.5 % growth in the preferred minimum acceptable monthly income and a 5.5% increase in the percentage of female daters who prefer home owners. However, for male local natives, the impacts of housing prices have different patterns. The only significant coefficient is for the regression where the outcome variable is preferred height range. A significantly negative coefficient indicates that male local natives respond to rising housing prices by narrowing down their preferred height range. A 1% increase in housing prices leads to a decrease of 0.006 centimeters in the preferred height range. Again, if the housing prices doubled, it would cause a decrease of 0.38 centimeters in the preferred height range. Overall, the SUR results provide two main observations. First, as local housing prices increase, female daters tend to have stronger preferences for males with high income, males who are wealthy, and males who are home owners. This impact is both statistically and economically significant. Second, male daters tend to narrow down their preferences on females height after they experience a positive housing-price shock, though the magnitude 10

including age, age squared, height, marital history, years of education, reported income (lower bound), reported income (upper bound), home ownership, reported fitness, car ownership, fertility history, weight, industry, self-ranked appearance, current condition of parents, number of siblings, smoking or not, drinking alcohol or not, ethnicity, and religious beliefs. Please refer to Appendix Table A1 for explanations of each variable.

11

of this impact is not very large. The results support gender-specific mating preferences, revealing that females care more about wealth and social economic status, while males care more about appearance.

4.2

Multinomial Logit Results

Table 4 presents the multinomial logit regression studying the impact of housing prices on the relative importance of housing preference and the other four preferences. We convert the continuous variables of preferences into dummy indicators as described in the previous section. Again, for each of the four preferences, there are four possible options for the joint preferences with housing preferences. Take the joint preferences of housing and income as an example: The online dater could have strong preferences regarding both housing and income (option 1), weak preferences regarding both housing and income (option 2), a strong preference regarding housing and a weak preference regarding income (option 3), or the reverse (option 4). In the regression in column 1, where the outcome variable is income preference, the coefficient on housing prices reports the impact of housing prices on the change in the decision between the last two options, where option 4 is taken as the baseline group. Therefore, a positive coefficient indicates that more people shift from choosing option 3 to option 4 as housing prices rise. Significant coefficients are reported for both male and female panels in column 4, where the joint preferences are housing and height. Interestingly, the signs of the coefficients are opposite for females and males. A significantly positive coefficient in Panel A suggests that more females shift from a a strong preference for height and a weak preference for housing to a strong preference for housing and a weak preference for height. In other words, female daters are more likely to compromise their height requirement for home owners when local housing prices rise. However, male daters are the reverse. A negative coefficient in Panel B, column 4 indicates that male daters are more likely to compromise on their preference for home owners for a narrower height range. The results are consistent with the findings in the SUR regressions that females have a stronger preference for home owners with an increase in local housing prices, while males have a stronger preference regarding the appearance of females, as measured by their preference on height. Furthermore, the multinomial logit regressions suggest that rises in housing prices not only reinforce gender-specific preferences but also affect the relative importance of preferences because people are more willing to compromise other preferences for their most desired ones.

12

4.3

Robustness Checks

We have conducted a series of robustness checks to make sure that the main results are consistent under alternative specifications. In the first robustness check, we test whether the multinomial logit estimation results are sensitive to using different sets of fixed effects. In the SUR estimations, we use the city by year fixed effects to control for the city-/time-varying unobserved factors. However, controlling such fine fixed effects will create difficulties in some MNL estimations due to the sparsity of the choice groups. Therefore, we use city and year fixed effects rather than city by year fixed effects in the MNL estimation. In this section, we show that our results are not sensitive to the use of alternative fixed-effect controls. In Table A2, we control the city by year fixed effects for the basic MNL estimations. The results in Columns 1 and 3 in Panel B are missing in this case. However, comparing the results in the other columns with Table 4 suggests that our results are robust to the lower level of fixed-effect controls. Another concern is regarding the definition of the binary preferences on income, education, age, and height. In the previous sections, we define a high-income preference as one’s stated preferred income being at least 5000 yuan. However, one potential challenge is that the economic heterogeneity among regions in China should be considered in the definition of high income. To address this concern, we compare the stated income preference with the average monthly income of a users city in the current registration year: It is 1 if his/her preferred minimum monthly income is above the average monthly income of their city and 0 otherwise. Regarding education preference, we define it as strong if the stated value is higher than the median value of the preferences specified by natives of the same gender. We similarly redefine the binary preference categories of age and height by comparing the age and height preference range with the median value of the same gender: The value is 1 if the stated preference is lower than the median value of the age (or height) preference range of the same gender. Table A3 reports the multinomial logit results using the alternative definitions. The results are largely consistent with the main results in Table 4. The only exception is that the significance level decreases for the relative odds of choosing home ownership and height in the male panel. However, the sign and magnitude remain similar. Lastly, one might worry that our results could be driven by the possibility that online daters strategically time themselves to join the dating website after observing the housing prices and thus that the groups of daters that we observe in different times are essentially different. To mitigate this concern, we regress the number of registrants in different months and cities on the lagged housing prices, after controlling for city-year fixed effects. The results in Table A4 suggest that housing prices do not affect the total number of male and female registrants on the website.

13

5 5.1

Discussions Heterogeneous Impacts by Home Ownership

In the main specification, we assume that local housing-price fluctuations serve as exogenous shocks to the perceived wealth of the local natives due to the high home ownership rate in China. In other words, local natives are very unlikely to be renters. They either own a house or live with their parents who own housing properties in the city. A natural question to ask is whether the effect of housing prices on mating preferences is more pronounced for home owners, i.e., those who own housing properties under their own name. In their profiles, online daters are asked about their home ownership status, so we know whether they live in their own house or they live with their parents.11 We define online daters as home owners if they report that they own a house (either with a mortgage or without a mortgage). We define people who choose other housing arrangements as non-home owners. As we are not sure about the home ownership status of users who choose not to reveal their information, we exclude them from the sample. We therefore interact home ownership with the housing-price measure and include them in the SUR and multinomial logit regressions. Table 5 reports the SUR regression results for the heterogeneous impacts of housing prices by home ownership. The other specifications are the same as in Table 3. Interestingly, we find a very positive and significant coefficient for the interaction term in Panel A. It suggests that local female home owners have a stronger preference for male home owners than for local non-home owners as housing prices rise. For local male daters, the interaction term has a significantly negative coefficient for the preference on age range, which suggests that male home owners are more likely to specify a narrower range for preferred age compared to non-home owners as local housing prices rise. The main effect of housing prices on preferred age is not significant, however. In addition, the main effect of housing prices on different dimensions of preferences is consistent with what is shown in Table 3. The results generally indicate that local home owners, who are likely to experience larger positive wealth shocks compared to non-home owners in housing booms, reinforce their preferred attributes more than non-home owners. This finding is consistent with our story that the impact of housing prices on mating preferences is likely to work through the wealth shock channel. Table 6 reports the multinomial logit regression results, including the interaction of home owners and housing prices. The main effects of housing prices are very similar to the results in Table 4. However, the substitution pattern does not differ by home ownership because 11

They can choose from the following categories: 1. own a house; 2. own a house (no mortgage); 3. own a house (with mortgage); 4. live with parents; 5. stay with relatives; 6. stay in rented apartment provided by company; 7. renter; 8. I will tell you later.

14

none of the coefficients for the interaction term is significant. The signs of the coefficients in column 4 are in the hypothesized directions, however, i.e., female home owners are more likely to trade their height requirements with home ownership compared to non-home owners, while male home owners are the reverse.

5.2

Heterogeneous Impacts by Age

As age is probably one of the most important demographic characteristics in the marriage market, we are interested in whether daters at different ages may respond to housing prices differently in terms of specifying their preferences. Therefore, we interact the dater’s age with the housing-price measure in both the SUR and multinomial logit regressions. Table 7 reports the SUR regression results after including the interaction term between age and housing prices. For female natives, as housing prices rise, older female daters significantly increase their preference for home owners compared to younger females. Otherwise, no heterogeneous effects are found in other preferred attributes. However, for male natives, as housing prices rise, they significantly increase their preferences in multi-dimensions: Older male daters increasingly prefer females who have higher income, own a house and have more years of education. They also seem to have a narrower range of preferred height. We do not have a well-established story about the gender-specific patterns of such heterogeneous effects by age. Intuitively, as age increases, females are more eager to get married and settle down. Therefore, the preference for home owners increases because home ownership provides females a sense of stability. In contrast, older males may have a better position in the marriage market due to wealth accumulation over the years; therefore, they are likely to have higher standards for their partners compared to younger males. Table 8 shows the multinomial logit results. As age increases, local female daters are more likely to compromise on the appearance of males (such as age and height) to trade for home ownership in a booming housing market, as shown by the significant coefficients in the interaction terms in columns 3 and 4. In contrast, male daters seem to be more willing to relax their preferences on home ownership to trade for higher-income females, as suggested by the marginally negative coefficient in column 1.

6

Conclusion

This paper investigates the impact of housing prices on mating preferences. Matching users stated preferences in an online mating website with prefecture city-level housing prices in China, we find that a positive wealth shock due to a booming housing market affects the mating preferences of local natives in gender-specific ways. As housing prices rise, female daters

15

have stronger preferences for wealthier males, as indicated by their specified preferences on monthly income and home ownership, while male daters narrow down their preferences for females’ appearance, as measured by their preferred ranges for females’ height. In addition, we find that housing prices affect the relative preferences for home ownership and appearance for both genders. As housing prices go up, females have a stronger preference on home ownership than on males’ appearance, while males are the opposite. To sum up, this paper conveys the findings that positive wealth shocks enhance daters preferences on their desired attributes, which are wealth and appearance for females and males, respectively. Moreover, positive wealth shocks enhance the relative importance of their desired attributes compared to other attributes. Taking one step further, it will be interesting to understand the impact of housing prices on matching outcomes if data are available. We leave this task for future research.

References [1] Mich`ele Belot and Marco Francesconi. Dating preferences and meeting opportunities in mate choice decisions. Journal of Human Resources, 48(2):474–508, 2013. [2] Martin Browning, Pierre-Andr´e Chiappori, and Yoram Weiss. Economics of the Family. Cambridge University Press, 2014. [3] Gustaf Bruze. Marriage choices of movie stars: Does spouses education matter? Marriage, 5(1), 2011. [4] Simon Chang and Xiaobo Zhang. Mating competition and entrepreneurship. Journal of Economic Behavior & Organization, 2015. [5] Pierre-Andr´e Chiappori, Sonia Oreffice, and Climent Quintana-Domeque. Fatter attraction: anthropometric and socioeconomic matching on the marriage market. Journal of Political Economy, 120(4):659–695, 2012. [6] Eugene Choo and Aloysius Siow. Who marries whom and why. Journal of political Economy, 114(1):175–201, 2006. [7] Yongheng Deng, Joseph Gyourko, and Jing Wu. Land and house price measurement in china. Technical report, National Bureau of Economic Research, 2012. [8] Yongheng Deng, Randall Morck, Jing Wu, and Bernard Yeung. Monetary and fiscal stimuli, ownership structure, and china’s housing market. Technical report, National Bureau of Economic Research, 2011. 16

[9] Lisa J Dettling and Melissa S Kearney. House prices and birth rates: The impact of the real estate market on the decision to have a baby. Journal of Public Economics, 110:82–100, 2014. [10] Paul W Eastwick and Eli J Finkel. Sex differences in mate preferences revisited: do people know what they initially desire in a romantic partner? Journal of personality and social psychology, 94(2):245, 2008. [11] Marcel Fafchamps and Agnes Quisumbing. Assets at marriage in rural ethiopia. Journal of Development economics, 77(1):1–25, 2005. [12] Raymond Fisman, Sheena S Iyengar, Emir Kamenica, and Itamar Simonson. Gender differences in mate selection: Evidence from a speed dating experiment. The Quarterly Journal of Economics, pages 673–697, 2006. [13] G¨ unter J Hitsch, Ali Horta¸csu, and Dan Ariely. Matching and sorting in online dating. The American Economic Review, pages 130–163, 2010. [14] Matthijs Kalmijn. Intermarriage and homogamy: Causes, patterns, trends. Annual review of sociology, pages 395–421, 1998. [15] DuEwa Kamara. The effect of the probability of marriage on housing demand for single women. Journal of housing economics, 3(4):296–311, 1994. [16] Robert Kurzban and Jason Weeden. Do advertised preferences predict the behavior of speed daters? Personal Relationships, 14(4):623–632, 2007. [17] Samantha Nelson, Lucy Delgadillo, and Jeffrey P Dew. Housing cost burden and marital satisfaction. Marriage & Family Review, 49(6):546–561, 2013. [18] David Ong and Jue Julie Wang. Income attraction: An online dating field experiment*. Journal of Economic Behavior & Organization, 2014. [19] Shang-Jin Wei and Xiaobo Zhang. Sex ratios, entrepreneurship, and economic growth in the peoples republic of china. Technical report, National Bureau of Economic Research, 2011. [20] Shang-Jin Wei, Xiaobo Zhang, and Yin Liu. Status competition and housing prices. Technical report, National Bureau of Economic Research, 2012. [21] Edward N Wolff. Recent trends in household wealth in the united states: Rising debt and the middle-class squeeze-an update to 2007. 2010.

17

[22] Jing Wu, Joseph Gyourko, and Yongheng Deng. Evaluating conditions in major chinese housing markets. Regional Science and Urban Economics, 42(3):531–543, 2012. [23] Yu Xie and Yongai Jin. Household wealth in china. Chinese sociological review, 47(3):203– 229, 2015.

18

Figure 1: Geographical Distribution of Users by Gender

(a) Female Users Across Cities

(b) Male Users Across Cities

19 GIS source: Michigan China Data Center.

Figure 2: Specified Preferences

(a) Preference on Housing by Home Ownership

(b) Preference on Housing by Age Group

20

(c) Preference on Income

(d) Preference on Education

21

(e) Preference on Age (in Range)

(f) Preference on Height (in Range) Source: authors’ calculation.

22

Figure 3: Availability of Housing Price Data

GIS source: Michigan China Data Center.

23

Figure 4: Housing Prices and Home Ownership Preference: Examples of Beijing and Shanghai

(a) Beijing

(b) Shanghai Note: The graphs above show the trends of housing prices and fraction of daters preferring home ownership for Beijing and Shanghai natives respectively. Monthly housing price is from Soufun Database, and the information about home ownership preference is obtained from our online dating database. For each city, the fraction of home ownership preferences by gender are calculated by averaging the dummy variable pref house, which indicates whether an individual prefers home owners or not (1 is yes and 0 otherwise), across the newly registered natives in every month.

24

Table 1: Descriptive Statistics: Individual Characteristics

25

Variable ownhouse

Description home ownership (1 is yes)

age

age (in years)

height

height (in cm)

never married before

1 is yes

divorced

1 is yes

education

years of education

incomel

lower bound of monthly income range (RMB)

incomeu

upper bound of monthly income range (RMB)

owncar

car ownership (1 is yes)

weight

weight (in kg)

N

Number of Observations

Female Natives 0.165 (0.371) 32.15 (9.316) 161.7 (4.468) 0.608 (0.488) 0.355 (0.479) 14.27 (1.896) 2599.3 (2324.7) 3999.8 (2905.2) 0.0897 (0.286) 52.99 (7.925) 60749

Male Natives 0.344 (0.475) 31.47 (8.536) 171.9 (6.262) 0.708 (0.455) 0.269 (0.444) 14.08 (1.998) 4090.4 (3685.2) 5792.7 (5215.2) 0.167 (0.373) 65.74 (11.02) 131701

Total 0.291 (0.454) 31.67 (8.782) 168.9 (7.427) 0.678 (0.467) 0.295 (0.456) 14.14 (1.970) 3644.3 (3406.1) 5256.4 (4718.2) 0.144 (0.351) 61.93 (11.75) 192450

Note: This table presents the summary statistics of some major variables by cohorts. The sample covers online dating users registered between 2012 and 2014. In our data, the natives refer to people whose residential city is the same as their stated home town. However, since a substantial proportion of people did not reveal their home city information, the number of natives is expected to be underestimated.

Table 2: Descriptive Statistics: Mating Preferences

26

Variable pref income

Description lower bound of acceptable income range (RMB/month)

pref house

stated preference on home ownership (1 is yes, 0 otherwise)

pref educ

lower bound of acceptable education level (years)

pref age upper

upper bound of preferred age (years)

pref age lower

lower bound of preferring age (years)

pref age range

age preference range (years)

pref height lower

lower bound of preferring height (cm)

pref height upper

upper bound of preferring height (cm)

pref height range

height preference range (cm)

N

Number of Observations

Female Natives 2843.2 (3920.3) 0.413 (0.492) 12.20 (3.403) 37.89 (9.939) 30.32 (8.735) 7.575 (3.225) 169.0 (6.026) 181.8 (5.306) 12.87 (5.539) 60749

Male Natives 675.0 (1888.4) 0.0602 (0.238) 10.45 (2.601) 31.31 (8.048) 24.33 (6.692) 6.981 (3.436) 157.2 (6.803) 172.8 (6.624) 15.59 (5.898) 131701

Total 1363.4 (2886.5) 0.172 (0.377) 11.01 (2.993) 33.40 (9.216) 26.23 (7.909) 7.169 (3.381) 161.0 (8.541) 175.7 (7.514) 14.73 (5.924) 192450

Note: The summary statistics in this table focus on the stated mating preference of different cohort groups. For the income preference, we define the income as 0 if one states ”any” or ”no specific requirement”. For the preference on education level, the lowest bound is set to be 9 in the cases that one replies ”any” or ”no specific requirement”, as we assume that people at least finish the 9-year compulsory education.

Table 3: Seemingly Unrelated Regression Estimations ln(pref income) Panel A: Female Natives Lag[ln(P )] Other Covariates N Panel B: Male Natives Lag[ln(P )] Other Covariates N

pref house

pref educ

pref age range

pref height range

(1)

(2)

(3)

(4)

(5)

0.585∗∗ (0.284) Yes 29303

0.079∗∗ (0.034) Yes 29303

0.222 (0.221) Yes 29303

-0.167 (0.221) Yes 29303

-0.364 (0.387) Yes 29303

0.249 (0.166) Yes 57232

0.018 (0.012) Yes 57232

0.143 (0.132) Yes 57232

-0.247 (0.158) Yes 57232

-0.553∗ (0.297) Yes 57232

Note: The regression results above are from the Seemingly Unrelated Regression (SUR) estimations. Standard errors are in parentheses. The dependent variables are defined in the text. As the lowest level of income preference is set at 0 if one does not state any specific preference, we replace it as 1 when we calculate the natural logarithms for this variable. The other control variables in the regressions include the detailed personal characteristics and city-year fixed effects. Standard errors are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

27

Table 4: Multinomial Logit Estimation

Panel A: Female Natives Lag[ln(P )] Other Covariates N Panel B: Male Natives Lag[ln(P )] Other Covariates N

Income

Education

Age

Height

(1)

(2)

(3)

(4)

0.265 (0.200) Yes 29406

0.243 (0.199) Yes 29406

0.029 (0.204) Yes 29406

0.504∗∗∗ (0.162) Yes 29406

-0.159 (0.314) Yes 57366

-0.123 (0.227) Yes 57366

-0.094 (0.189) Yes 57366

-0.601∗ (0.331) Yes 57366

Note: In this table, we report the multinomial logit (MNL) estimations that examine the potential substitution between the preference on home ownership and the other four preferences. We convert the continuous variables of preferences into dummy indicators as described in the text. The above multinomial Logit estimations use the joint preferences of home ownership preference and the other four preferences as dependent variables, respectively. The baseline groups are set as (0, 1), that is, no stated preference on home ownership but strong preference on the other attribute. The estimated coefficients on group (1, 0) for each MNL regression are reported. The city and year fixed effects, as well as city-level variables, including GDP per capita, population, income per capita, and cohort specific sex ratio are controlled in the MNL estimations. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

28

Table 5: SUR Estimations: Heterogeneous Effects by Home Ownership ln(pref income) Panel A: Female Natives Lag[ln(P )] Lag[ln(P )] × own Other Covariates N Panel B: Male Natives Lag[ln(P )] Lag[ln(P )] × own Other Covariates N

pref house

pref educ

pref age range

pref height range

(1)

(2)

(3)

(4)

(5)

0.658∗ (0.348) -0.175 (0.131) Yes 17531

0.079∗ (0.043) 0.045∗∗∗ (0.016) Yes 17531

0.120 (0.275) -0.116 (0.103) Yes 17531

-0.327 (0.275) -0.073 (0.104) Yes 17531

-0.460 (0.475) 0.035 (0.179) Yes 17531

0.261 (0.214) 0.003 (0.071) Yes 35543

0.017 (0.016) 0.004 (0.005) Yes 35543

0.129 (0.173) 0.033 (0.058) Yes 35543

-0.089 (0.208) -0.156∗∗ (0.070) Yes 35543

-0.708∗ (0.383) -0.119 (0.128) Yes 35543

Note: The home ownership dummy that interacted with the log of housing price in the regressions is defined to be 1 if one specifies he/she owns a property (with mortgage or without mortgage), and 0 otherwise. In these regressions, we do not include the observations that provide no information on the ownership of a property. The other control variables in the regressions include detailed personal characteristics and city-year fixed effects. Standard errors are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

29

Table 6: Multinomial Logit Estimation: Heterogeneous Effects by Home Ownership

Panel A: Female Natives Lag[ln(P )] Lag[ln(P )] × own Other Covariates N Panel B: Male Natives Lag[ln(P )] Lag[ln(P )] × own Other Covariates N

Income

Education

Age

Height

(1)

(2)

(3)

(4)

0.240 (0.323) 0.158 (0.151) Yes 17611

0.273 (0.266) 0.220 (0.164) Yes 17611

-0.061 (0.359) 0.137 (0.084) Yes 17611

0.440∗ (0.245) 0.172∗ (0.096) Yes 17611

-0.254 (0.375) -0.168 (0.148) Yes 35639

-0.428 (0.308) 0.200 (0.122) Yes 35639

-0.444∗ (0.252) -0.018 (0.156) Yes 35639

-1.175∗∗∗ (0.373) -0.064 (0.157) Yes 35639

Note: The home ownership dummy that interacted with the log of housing price in the regressions is defined to be 1 if one specifies he/she owns a property (with mortgage or without mortgage), and 0 otherwise. In these regressions, we do not include the observations that provide no information on the ownership of a property. The city and year fixed effects, as well as city-level variables, including GDP per capita, population, income per capita and cohort specific sex ratio are controlled in the MNL estimations. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

30

Table 7: SUR Estimations: Heterogeneous Effects by Age ln(pref income) Panel A: Female Natives Lag[ln(P )] Lag[ln(P )] × age Other Covariates N Panel B: Male Natives Lag[ln(P )] Lag[ln(P )] × age Other Covariates N

pref house

pref educ

pref age range

pref height range

(1)

(2)

(3)

(4)

(5)

0.733∗∗ (0.329) -0.005 (0.005) Yes 29303

0.029 (0.040) 0.002∗∗ (0.001) Yes 29303

0.362 (0.256) -0.004 (0.004) Yes 29303

-0.226 (0.255) 0.002 (0.004) Yes 29303

-0.539 (0.447) 0.005 (0.007) Yes 29303

-0.044 (0.194) 0.009∗∗∗ (0.003) Yes 57232

-0.010 (0.014) 0.001∗∗∗ (0.000) Yes 57232

-0.030 (0.154) 0.005∗∗ (0.002) Yes 57232

-0.260 (0.186) 0.000 (0.003) Yes 57232

-0.226 (0.349) -0.010∗ (0.006) Yes 57232

Note: The regression results above are from the Seemingly Unrelated Regression (SUR) estimations. Standard errors are in parentheses. The dependent variables are defined in the text. As the lowest level of income preference is set at 0 if one does not state any specific preference, we replace it as 1 when we calculate the natural logarithms for this variable. The other control variables in the regressions include detailed personal characteristics and city-year fixed effects. Standard errors are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

31

Table 8: Multinomial Logit Estimation: Heterogeneous Effects by Age

Panel A: Female Natives Lag[ln(P )] Lag[ln(P )] × age Other Covariates N Panel B: Male Natives Lag[ln(P )] Lag[ln(P )] × age Other Covariates N

Income

Education

Age

Height

(1)

(2)

(3)

(4)

0.123 (0.425) 0.004 (0.012) Yes 29406

0.060 (0.417) 0.004 (0.011) Yes 29406

-0.285 (0.226) 0.010∗∗ (0.004) Yes 29406

0.179 (0.202) 0.010∗∗ (0.004) Yes 29406

0.246 (0.472) -0.014∗ (0.007) Yes 57366

-0.256 (0.249) 0.001 (0.004) Yes 57366

-0.334 (0.253) 0.006 (0.005) Yes 57366

-0.558 (0.373) -0.001 (0.005) Yes 57366

Note: In this table, we report the multinomial logit (MNL) estimations that examine the potential substitution between the preference on home ownership and the other four preferences. We convert the continuous variables of preferences into dummy indicators as described in the text. The above multinomial Logit estimations use the joint preference of home preference and the other four preferences as dependent variables, respectively. The baseline groups are set as (0, 1), that is, no stated preference on home ownership but strong preference on the other attribute. The estimated coefficients on group (1, 0) for each MNL regression are reported. The city and year fixed effects, as well as city-level variables, including GDP per capita, population, income per capita and cohort specific sex ratio are controlled in the MNL estimations. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

32

Appendix Table A1: Variable Definition Personal Attributes

Description

Personality Type Resident City Age Height Marital History Education

16 types of personalities Current resident city (prefecture level) Age (in years) Height (in cm) Unmarried, Divorced, Widowed {Education level (junior high/senior high/senior high or below/ vocational school/vocational college/university/bachelor degree/master degree/phd/postdoc) Monthly Income expressed in range More than 60 categories of occupation Home owner/Home owner (without mortgage)/Home owner (with mortgage)/Consider home purchase whenever necessary/Live with parents/Stay in coporate subsidized apartment/Live with relatives or friends/Renter) Hometown (prefecture level) Alma mater Self described fitness level (from very thin to very strong) Major in college (30 categories) 12 Animal Zodiac Signs Own a car/not own a car/use a car provided by corporate/consider car purchase whenever necessary 12 zodiac signs indicating birth month No kids/Have kids who stay with me/Have kids who sometime stay with me/Have kids who don’t stay with me Reported weight in kg Mastered languages and dialect AB/A/B/O/Others From 0-10 (higher score indicates better appearance) Han or other minorities Taoism/Buddism/Islam/Christianity/Confusianism/Catholicism/No religious belief/Hinduism/Judaism

Monthly Income Occupation Living Arrangement

33

Hometown Alma mater Fitness Major Chinese Animal Zodiac Car Ownership Zodiac Sign Fertility history Weight Language Proficiency Blood type Self-ranked appearance Ethnicity Religious belief

Table A2: Alternative Control for Fixed Effects in the MNL Estimations

Panel A: Female Natives Lag[ln(P )] Other Covariates N Panel B: Male Natives Lag[ln(P )]

Income

Education

Age

Height

(1)

(2)

(3)

(4)

0.233 (0.214) Yes 29406

0.260 (0.217) Yes 29406

0.201 (0.250) Yes 29406

0.617∗∗∗ (0.204) Yes 29406

-0.183 (0.293) Yes 57366

Other Covariates N

-0.397 (0.377) Yes 57366

Note: In this table, the multinomial logit estimations are conducted by controlling different sets of fixed effects for the female group and male group respectively. In both panels, we control the city-year fixed effects. However, as some choice outcomes are too sparse in the MNL regression for the male group after controlling these fixed effects, we are not able to obtain the point estimations for the odds ratio in some regressions for the male group. The purpose of this table is to provide evidence that our results are not sensitive to the alternative control for fixed effects. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

34

Table A3: Alternative Definition of Income, Education, Age Preference, and Height Preference in the MNL Estimations

Panel A: Female Natives Lag[ln(P )] Other Covariates N Panel B: Male Natives Lag[ln(P )] Other Covariates N

Income

Education

Age

Height

(1)

(2)

(3)

(4)

-0.010 (0.314) Yes 29406

0.207 (0.189) Yes 29406

0.056 (0.190) Yes 29406

0.437∗∗ (0.185) Yes 29406

-0.755 (0.574) Yes 57366

0.050 (0.232) Yes 57366

-0.012 (0.172) Yes 57366

-0.397 (0.290) Yes 57366

Note: In these robustness checks of MNL estimations, we use different definitions of high preferences on income, education, age, and height respectively. Specifically, for the income preference binary indicator, we define it is 1 if a people’s stated preference on income is higher than the average monthly income of his/her city at that year. Considering the spatial heterogeneity of education level, we define a people has relatively high preference on education if the stated value is higher than the median level of his/her counterparts (the same gender of natives in the same city). This method is also applied for the new definition of strict preference on age and height, by comparing the age preference range and height preference range with the median value of the same cohort in the same city. The city and year fixed effects, as well as city-level variables, including the GDP per capita, population, income per capita and cohort specific sex ratio are controlled in the MNL estimations. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

35

Table A4: Falsification: Number of Website Registrants and Housing Prices

Lag[ln(P )] City by year fixed effects N

Male

Female

(1) -0.011 (0.149) Yes 3068

(2) 0.108 (0.112) Yes 2997

Note: The dependent variable is monthly number of female and male registrants from each city. We control the city by year fixed-effects in the Poisson regressions. Standard errors clustered at the city level are reported in parentheses. ∗ Significant at the 10 percent level. ∗∗ Significant at the 5 percent level. ∗∗∗ Significant at the 1 percent level.

36