JUE Insight: COVID-19 and Household Preference for Urban Density in China.

This paper investigates the effect of COVID-19 on both housing prices and housing price gradients in China using transaction level data from 60 Chinese cities. After using a difference-in-differences (DID) specification to disentangle the confounding effects of China's annual Spring Festival, we find that housing prices decreased by two percent immediately after the COVID-19 outbreak but gradually recovered by September 2020. Moreover, our findings suggest that COVID-19 flattens the horizontal housing price gradient, reduces the price premium for living in tall buildings, and changes the vertical gradient within residential buildings. This is likely explained by the changing household preferences towards low-density areas associated with lower infection risk.

Introduction

While urban density is central to agglomeration economies and economic growth, COVID-19 may have changed household preferences for it. Current studies have shown that the bid-rent curve became flatter during the pandemic (Brueckner et al., 2021; Cheung et al., 2021; Gupta et al., 2021; Liu and Su, 2021; Rosenthal et al., 2022), suggesting that households are increasingly interested in properties farther away from the city center. Unlike its effect on the horizontal price gradient, the potential effect of COVID-19 on the vertical pattern of urban development has received little attention. To our knowledge, this is the first paper to fill in the gap and study the effect of COVID-19 on both the horizontal and vertical price gradients.

Using unit-level housing transactions from the largest online real estate agency in China, we analyze the impact of the pandemic on housing prices, and on the horizontal and vertical housing price gradients. We adopt a difference-in-differences (DID) specification to disentangle the confounding effect of the Chinese lunar New Year, also known as the Spring Festival, which was held very close to the outbreak of the pandemic. This DID setting compares the outcome variables in 2020 with their counterparts in the same lunar calendar period in 2019. In addition to the DID strategy, an event study is used to explore the dynamic effects of COVID-19 and illustrate pre-trends prior to the outbreak of the pandemic.

We find that housing prices decreased by two percent immediately after the outbreak of the pandemic but had returned to pre-pandemic levels by the beginning of September 2020. Residential units near hospitals and locations where inhabitants contracted COVID-19, and in cities with strict social-distancing policies all experienced greater drops in prices. COVID-19 also changed both the horizontal and vertical housing price gradients. Results suggest that COVID-19 flattened the housing price gradient with respect to distance, as the prices of apartments closer to city centers experienced greater decreases. The virus has also affected the preference over vertical densities measured by building heights and floor levels within buildings, even after controlling for the changes in the horizontal price gradient. Specifically, we find that apartments in taller residential buildings suffered from a greater price decline during the pandemic and housing prices of apartments on higher floors experienced larger increases compared to those on lower floors. All these findings are consistent with the explanation that COVID-19 reduces household preferences for living in denser areas with higher infection risk.

This paper contributes to the current literature in several ways. First, this paper is the first to analyze COVID-19’s effect on both horizontal and vertical gradients. Current studies on COVID-19 have predominantly focused on how the pandemic affected the horizontal housing price gradient in cities (Brueckner et al., 2021; Gupta et al., 2021; Liu and Su, 2021; Rosenthal et al., 2022). However, horizontal density does not necessarily reflect the variation of vertical density: the correlation coefficient between distances to city centers and building heights is only 0.06 in Chinese cities. Integrating the analysis of both gradients in the same framework enables us to better understand how COVID-19 has changed the internal structure of cities.

Second, this paper contributes to the literature on tall buildings. There is a growing literature on the economics of skyscrapers (Ahlfeldt and McMillen, 2018; Ahlfeldt and Barr, 2020, 2022; Albouy et al., 2020). However, none of these studies evaluates whether COVID-19 has changed household preferences for tall buildings. This paper also extends the literature on the bid-rent curve within tall buildings. Different from Liu et al. (2018, 2020), who study commercial properties, this paper provides new empirical evidence regarding residential homes and explores whether the vertical pattern within buildings has changed after COVID-19.

Lastly, this paper complements existing studies on pandemic and housing prices by analyzing unit-level housing transaction records in the 60 largest Chinese cities. Most existing studies on this topic use monthly community-level data or only focus on one city (Cheung et al., 2021; Liu and Tang, 2021; Qian et al., 2021). The data used in this paper is more granular and representative of the Chinese housing market.

The remainder of the paper is structured as follows. The next section discusses potential mechanisms through which COVID-19 may have affected horizontal and vertical price gradients. Section 3 introduces the data and Section 4 describes identification strategies. Empirical results are reported in Section 5. We conclude the paper in Section 6.

Why COVID-19 could affect housing price gradients

The rapid spread of COVID-19, especially in crowded indoor spaces, has greatly raised public awareness and caused people to reconsider the benefits and costs of living in dense areas (Rosenthal et al., 2022). Crowded areas where various amenities are available may lose their appeal due to their perceived high infection risk. Meanwhile, the increasing availability of work-from-home could greatly reduce commuting costs and change households’ valuation of urban density. While there is still no consensus on how COVID-19 has affected the housing price gradient in general, we would like to discuss potential mechanisms through which it may have changed household preferences for urban densities and thus affect different types of housing price gradients as follows.

First, COVID-19 may reduce household preference for living close to city centers. The negative relationship between housing prices and the distance to city centers has long been recognized (Alonso, 1964; Mills, 1967; Muth, 1969; Duranton and Puga, 2015; Albouy et al., 2018; Huang et al., 2018). The downward sloping bid-rent curve reflects the increased willingness to bid for houses close to city centers where job opportunities are abundant. This relationship, however, may have been weakened due to the rise of remote working after the outbreak of the pandemic. Household preference for living near city centers may have further decreased if the perceived infection risk in city centers is higher.

Second, COVID-19 may have reduced household preferences for density based on building height. In addition to the distance to city centers, building height is an alternative measure for urban density (Albouy et al., 2020; Ahlfeldt and Barr, 2022). Taller buildings are perceived as having higher density as they accommodate more people, that is, one is likely to encounter more neighbors while waiting for elevators or jogging in the playground. The increased infection risk associated with COVID-19 is likely to decrease households’ willingness to pay for homes in taller residential buildings.

Lastly, the COVID-19 pandemic may have changed the bid-rent curve within residential buildings. The housing price across different floors reflects the trade-off between vertical transport costs and the amenities that vary with greater building height (Liu et al., 2018). While amenities such as better access to sunshine, fresh air, and scenic views would increase the price premium for houses on high floors, the increasing transport cost to reach these houses would make them less attractive. The effect of the pandemic on the vertical gradient within buildings represents the changes in preferences for those amenities and transport costs. During the pandemic, spending more time at home may have strengthened the value of scenic views at high floors. Meanwhile, households decreased outdoor activities which thus weakened the impact of transport costs. In addition, apartments on the ground floor have more foot traffic, as all residents enter there. The higher infection risk due to foot traffic may have decreased household valuations of apartments on low floors. Therefore, we expect that households may increase their willingness to pay for apartments on high floors.

In addition to building height, floor levels, and the distance to city centers, COVID-19 may also have changed households’ valuation of other local amenities such as hospitals, shopping malls, and museums. If the distribution of these amenities is correlated with our density measures, changes in households’ valuation of these amenities may also have caused changes in horizontal and vertical gradients. In this sense, our density measures capture amenities that are associated with urban density. While this paper focuses on urban density and involves analysis of hospitals, we leave future work to identify COVID-19’s effect on households’ valuation of other amenities.

Data

To empirically test our research question, we combine three groups of datasets from different sources.1 First, the second-hand housing transaction data is obtained from Beike Zhaofang, the largest online real estate agency in China. Each transaction record lists detailed information on the characteristics of the sold housing unit: transaction date, transaction price, unit attributes (e.g., size, floor level, style, and unit orientation), the name of the development project (xiao qu) where the unit is located, and the project's geographic location (the latitude and longitude of the project's centroid). There are 67,032 projects in our sample. A development project in a Chinese city is a small geographic unit. It is usually surrounded by walls or fences, and residents share one or several gates to enter or exit the project. A development project usually contains hundreds or thousands of housing units, and the number of residential buildings in a project ranges from several to dozens.

Our sample period includes the first 35 weeks in 2020 and the 35 weeks during the same lunar calendar period in 2019. Specifically, we include all sales transactions made between January 1 and September 1, 2020, as well as those between January 15 and September 16, 2019. Our sample includes the top 60 cities with the most observations.2 Most cities in our sample are first-tier cities, second-tier cities, and provincial capital cities.

Multiple variables measuring densities from different perspectives are constructed as follows. First, we calculate the distance to city centers from the geocodes of city centers and project centroids.3 We then use the total number of floors of the building where the housing unit is located as a proxy for the building height. In addition to these two density measures, we also consider the floor number within a residential building. However, the dataset does not report the exact floor number that a unit is located on. Instead, a unit is classified as located on either a low, middle, or high floor within the building.4

The second dataset includes the locations of confirmed COVID-19 cases and hospitals to proxy for potential infection risk. The locations of confirmed COVID-19 cases are collected from two websites which assembled daily confirmed cases and their geocoded locations from government news sources.5 Hospital addresses are gathered from the National Medical Insurance Administration and transformed into geocoded coordinates using services provided by Baidu Maps.

Following Qiu et al. (2020), the last dataset reports whether a city had adopted project-level outdoor activity restrictions. In cities that adopted these restrictions, household activities are greatly restricted to their projects. A common practice in these cities is that only one person in each household is allowed to go out every two days to purchase necessities, and permits are required when a resident exits or reenters the project. In our sample, 24 out of 60 cities adopted these strict social-distancing policies.6

Empirical strategies

As the COVID-19 outbreak (Wuhan lockdown on January 23, 2020) occurred right before the 2020 Spring Festival (January 25, 2020), this paper employs a DID specification to remove the festival's potential confounding effect. The Spring Festival is the most important event in China during which households are likely to consider where to live and work in the following year (Meng et al., 2021). We define the treatment group as apartments transacted during the 35 weeks between January 1, 2020 and September 1, 2020. The control group is defined as apartments transacted during the same lunar calendar period in 2019: the 35 weeks between January 15, 2019 and September 16, 2019. Both groups cover three weeks before and 31 weeks after the Spring Festival week.

Note that this DID specification is different from the traditional setting where the treatment and control groups experience contemporaneous changes. As COVID-19 affects all cities simultaneously, it is hard to find cities whose housing markets have not suffered from the pandemic. This leads us to use the performance of the housing market in 2019 as the control group. The implicit assumption is that the performance of the housing market in 2020 would be parallel to that in 2019 if the COVID-19 pandemic had not occurred. Nevertheless, the spirit of our specification is the same as the traditional DID method, as the identification comes from the differences in housing price (gradients) changes before and after the 2019 and 2020 Spring Festivals.7

The effect of COVID-19 on housing prices

The regression model is a standard hedonic model with a DID specification:where the dependent variable is the logarithm of the housing price per square meter for apartment in development project (xiao qu) on transaction date . is a constant term. is a dummy variable which equals one for observations in the treatment group and zero for observations in the control group. The dummy variable equals one if transaction date is after the 2019 or 2020 Spring Festival and zero otherwise. The coefficient () of the interaction term represents the difference between the change of housing prices before and after the 2020 Spring Festival compared to their counterparts during the same lunar calendar period in 2019. As the 2020 Spring Festival began only two days after the Wuhan lockdown, the coefficient () of the interaction term also captures the effect of COVID-19 on housing prices.8

To account for other factors that affect housing prices, we include a series of unit characteristics, denoted by in Eq. (1). The unit attributes include the logarithm of property size in square meters, the floor level (middle floors, high floors with elevators, and high floors without elevators, compared to low floors), the orientation of the housing unit (facing north or not), and the style type (simple, fancy, or other, compared to unfurnished).9

We also include project fixed effects to account for neighborhood amenities, such as school quality, access to transit, and distance to employment centers. As the time span of our study period is less than two years, these amenities are not likely to have changed during the sample period. The inclusion of project fixed effects guarantees the inclusion of city fixed effects, as a development project in China never crosses city borders.

A series of week-pair fixed effects, denoted by, are used to capture the common week-specific factors in both 2019 and 2020 that may influence the housing market. denotes the week “distance” to the Spring Festival of each year for units transacted on date , which ranges from -3 to 31. For example, () equals one when an apartment was transacted during the second week before (after) the 2019 or 2020 Spring Festival. The week-pair fixed effect can account for cyclical confounding factors that are sensitive to the lunar calendar, such as the massive migration flow during the Spring Festival and its subsequent effect on housing purchase decisions. The error term is denoted by .

The effect of COVID-19 on housing price gradients

To identify the effect of COVID-19 on housing price gradients, we apply the following triple difference specification based on Eq. (1):

For unit , takes three groups of values. The first group measures the horizontal density: the logarithm of the straight-line distance between apartment and the city center.10 The second group captures the vertical density measured by the logarithm of building height (i.e., by the total number of floors) of a residential building where apartment is located. The last group is a set of dummies indicating the floor level where unit is located. The effect of COVID-19 on housing price gradients is captured by , the coefficient of the triple interaction term. When denotes the distance to the city center, a positive implies that COVID-19 causes households to prefer homes in suburban areas to units near city centers. Similarly, the estimate of should be negative when COVID-19 decreases household willingness to pay for apartments in taller residential buildings when represents the building height. Other variables are similarly defined as in Eq. (1).

As mentioned earlier, distance to city centers and building heights are not highly correlated in Chinese cities. Nevertheless, the distance to city centers and its interactions with the DID variables are controlled when we evaluate the effect of COVID-19 on the vertical gradient.

The dynamic effect from an event study

We rely on an event study to illustrate the dynamic effect of COVID-19 on the housing market. Another advantage of an event study is the capability to show pre-trends. To implement the event-study analysis, a series of dummies () representing the weeks before and after the COVID-19 outbreak is included in the following regression:where denotes the COVID-19 outbreak week. is an assignment function which equals one if is true and otherwise equals zero. All other variables are similarly defined as before. To keep the sample consistent in the main analysis, we first choose the study window as three weeks before and 31 weeks after the COVID-19 outbreak week, i.e., , , and As a robustness check, we then allow for longer periods by including eight additional weeks before the Spring Festival. To avoid perfect collinearity, the week before the COVID-19 outbreak week is set as the base period, i.e., is normalized to zero. The coefficient thus represents the price change in week with respect to the price in the base period. Eq. (3) identifies the dynamic effect of COVID-19 on housing prices, and a similar specification is applied to study the dynamic effect of COVID-19 on housing price gradients.

Results

Table 1 reports the estimates for the effect of COVID-19 on housing prices obtained from Eq. (1). Standard errors are clustered at the project level. All control variables are included in the regression but omitted in Table 1 to save space. The empirical result in column (1) suggests that COVID-19 significantly decreases average housing prices by 0.7%. This estimate is qualitatively similar to existing studies that also identify a negative effect of COVID-19 on housing prices in China (Cheung et al., 2021; Liu and Tang, 2021; Qian et al., 2021).

Table 1

The impact of Covid-19 on average housing prices.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	All projects	Projects with confirmed Covid-19 cases	Projects without confirmed Covid-19 cases	Projects close to hospitals	Projects far from hospitals	Strict social-distancing policies	Less strict social-distancing policies
Year 2020 (Treat)	0.023***	0.036***	0.022***	0.016***	0.028***	0.015***	0.027***
	(0.001)	(0.004)	(0.001)	(0.002)	(0.002)	(0.002)	(0.002)
After SF (SF_t)	0.038***	0.041***	0.038***	0.034***	0.041***	0.033***	0.040***
	(0.001)	(0.004)	(0.002)	(0.002)	(0.002)	(0.003)	(0.002)
Treat×SF_t	-0.007***	-0.018***	-0.005***	-0.010***	-0.006***	-0.009***	-0.005***
	(0.001)	(0.003)	(0.001)	(0.002)	(0.002)	(0.002)	(0.001)
Observations	747117	111521	614623	265390	481727	266745	480372
R-squared	0.979	0.980	0.979	0.979	0.979	0.951	0.983
Project FE	YES	YES	YES	YES	YES	YES	YES
Week pair FE	YES	YES	YES	YES	YES	YES	YES

Notes: All regressions follow Eq. (1). Housing attributes are controlled but not displayed. represents whether the transaction date is after the Spring Festival in 2019 or 2020. denotes whether a transaction is in year 2020. The results in column (1) use all observations in the sample. In column (2), we define a project as having confirmed cases when there are confirmed Covid-19 cases within five hundred meters of the project centroid by April 8, 2020, when Wuhan reopened the city. Otherwise, projects are defined as not having confirmed cases in column (3). The results in columns (2) and (3) exclude four cities (Huhhot, Langfang, Shaoxing, and Shijiazhuang), because our dataset does not report the geocodes of confirmed Covid-19 cases in these four cities. A project is considered close to hospitals and included in column (4) if there is at least one hospital within 500 meters of the project centroid. Otherwise, a project is defined as far from hospitals and included in column (5). Column (6) includes cities that adopted stringent outdoor activity restrictions where residents were confined within projects. The other cities that did not adopt such policies are displayed in column (7). Standard errors in parenthesis are clustered at the project level. * indicates significance at the 10% level, ** indicates significance at the 5% level, and *** indicates significance at the 1% level.

The dynamic effect of COVID-19 on average housing prices is illustrated in Online Appendix Fig. A2(a). The figure only plots the coefficient estimates () of the week dummies in Eq. (3) and their corresponding 95% confidence intervals. We clearly observe a sharp two-percent decrease of housing prices immediately after the COVID-19 outbreak and a gradual recovery afterwards. This finding resembles that of Liu and Tang (2021), who also observed a rebound of housing prices in June 2020 following a housing price decline. By the end of our sample period, the average housing price had returned to pre-pandemic levels. There are no obvious pre-trends in the event-study graph, suggesting that the parallel trends assumption of the DID is likely to hold.11

Columns (2)–(7) of Table 1 explore potential heterogenous effects. Columns (2) and (3) compare the effect of COVID-19 for projects with and without confirmed COVID-19 cases. A project is defined as having confirmed cases if there existed at least one confirmed COVID-19 case within 500 m of the project centroid by April 8, 2020.12 The estimates in columns (2) and (3) indicate that projects with confirmed cases experienced greater decreases in housing prices after the COVID-19 outbreak, consistent with the finding in Liu and Tang (2021).

To formally test whether the difference of the estimates in columns (2) and (3) are statistically significant, we use a triple interaction term by interacting the DID variables with a new dummy variable representing whether a project had COVID-19 cases. The full results are presented in column (1) of Online Appendix Table A3. The coefficient of the additional interaction term indicates that the price decrease of homes with COVID-19 cases was significantly greater than that for those without COVID-19 cases.

The second heterogeneity analysis compares the effect of COVID-19 on the prices of projects close to and far away from hospitals. A project is considered close to hospitals if there exists at least one hospital within 500 m of the project centroid. As the proximity to hospitals could be correlated with a higher COVID-19 infection risk, home buyers may have paid a premium to live further away from hospitals.13 The estimates in columns (4) and (5) support this supposition, as projects close to hospitals experienced a larger decrease in housing prices. However, this difference is not statistically significant (see column (2) of Online Appendix Table A3).

The last heterogeneity analysis compares the effect of COVID-19 in cities with different social-distancing policies. According to Qiu et al. (2020), cities with strict social-distancing policies have stringent outdoor activity restrictions that confine residents to their projects. Empirical results in columns (6) and (7) confirm that COVID-19’s negative effect on housing prices is significantly greater in cities with more stringent social-distancing policies (see column (3) of Online Appendix Table A3).

The effect of COVID-19 on the housing price gradient with respect to distance

We estimate Eq. (2) to identify the effect of COVID-19 on the housing price gradient with respect to distance. In column (1) of Table 2 , the positive coefficient on the interaction term between the distance variable and the DID estimator implies that COVID-19 flattens the housing price gradient. This finding is consistent with Cheung et al. (2021), who find that the price premium of high-density areas in Wuhan decreased after the COVID-19 outbreak. The corresponding event-study result in Online Appendix Fig. A3 indicates that household preferences for suburban homes gradually emerged after the COVID-19 outbreak, and this change was not driven by pre-trends. It also suggests that the change in preferences persisted to the end of the sample period.

Table 2

The impact of Covid-19 on the housing price gradient with respect to distance.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	All projects	Projects with confirmed Covid-19 cases	Projects without confirmed Covid-19 cases	Projects close to hospitals	Projects far from hospitals	Strict social-distancing policies	Less strict social-distancing policies
Year 2020 (Treat)	0.044***	0.041***	0.041***	0.033***	0.063***	0.028***	0.054***
	(0.004)	(0.011)	(0.004)	(0.005)	(0.006)	(0.006)	(0.005)
After SF (SF_t)	0.050***	0.040***	0.051***	0.041***	0.059***	0.034***	0.059***
	(0.003)	(0.009)	(0.004)	(0.004)	(0.005)	(0.005)	(0.004)
Treat×SF_t	-0.020***	-0.031***	-0.016***	-0.022***	-0.020***	-0.007	-0.025***
	(0.004)	(0.010)	(0.004)	(0.005)	(0.005)	(0.006)	(0.005)
Treat×ln(dist)	-0.010***	-0.002	-0.009***	-0.009***	-0.015***	-0.007***	-0.012***
	(0.002)	(0.005)	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)
SF_t×ln(dist)	-0.005***	0.001	-0.006***	-0.004**	-0.008***	-0.001	-0.009***
	(0.001)	(0.004)	(0.001)	(0.002)	(0.002)	(0.002)	(0.002)
Treat×SF_t×ln(dist)	0.006***	0.007	0.005***	0.007***	0.006***	-0.001	0.009***
	(0.002)	(0.005)	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)
Observations	747117	111521	614623	265390	481727	266745	480372
R-squared	0.979	0.980	0.979	0.979	0.979	0.951	0.983
Project FE	YES	YES	YES	YES	YES	YES	YES
Week pair FE	YES	YES	YES	YES	YES	YES	YES

Notes: All regressions follow Eq. (2). Housing attributes are controlled but not displayed. represents whether the transaction date is after the Spring Festival in 2019 or 2020. denotes whether a transaction is in year 2020. The results in column (1) use all observations in the sample. In column (2), we define a project as having confirmed cases when there are confirmed COVID-19 cases within five hundred meters of the project centroid by April 8, 2020, when Wuhan reopened the city. Otherwise, projects are defined as not having confirmed cases in column (3). The results in columns (2) and (3) exclude four cities (Huhhot, Langfang, Shaoxing, and Shijiazhuang), because our dataset does not report the geocodes of confirmed Covid-19 cases in these four cities. A project is considered close to hospitals and included in column (4) if there is at least one hospital within 500 meters of the project centroid. Otherwise, a project is defined as far from hospitals and included in column (5). Column (6) includes cities that adopted stringent outdoor activity restrictions where residents were confined within projects. The other cities that did not adopt such policies are displayed in column (7). Standard errors in parenthesis are clustered at the project level. * indicates significance at the 10% level, ** indicates significance at the 5% level, and *** indicates significance at the 1% level.

Columns (2)–(7) of Table 2 report results for potential heterogeneities. From column (2), we find that COVID-19 had a positive effect on the slope of the housing price gradient of projects with confirmed COVID-19 cases. Due to the smaller sample size and larger standard deviations, this effect is not statistically significant. The effect for projects without confirmed COVID-19 cases in column (3) is significant and slightly lesser than the estimate in column (2). However, Online Appendix Table A4 shows that the difference between these two coefficients is not statistically significant.

As shown in columns (4) and (5), the effect of COVID-19 on the slope of the housing price gradient is significantly positive and similar in magnitude for projects close to and far away from hospitals. Thus, the effect of COVID-19 on the horizontal housing price gradient does not change with the proximity of confirmed COVID-19 cases or hospitals. While in Section 4.1 we find households display different levels of willingness to pay for homes with different levels of infection risk, we do not observe their preferences related to housing location changes across these groups of properties. One possible explanation is that people change their preferences for urban density independent of the risks associated with proximity to COVID-19 cases or hospitals.

However, we find heterogeneous effects in cities with different social-distancing policies. According to the results in columns (6) and (7), the effect of COVID-19 on the horizontal gradient is close to zero in cities with strict social-distancing policies, while the effect in cities with less strict social-distancing policies is significantly positive. This significant heterogeneity may be explained by the idea that households in cities with strict social-distancing policies had to stay at home or were confined within their projects. As a result, the infection risk mainly came from contact with other people within the project. The benefits of living in the suburban areas of cities with strict social-distancing policies would be arguably lower than in cities with laxer policies. Therefore, households in these cities may have had fewer incentives to relocate to suburbs.

To further explore the changes in the slope of the housing price gradient, we discretize the distance variable () and report the estimates for the interaction terms between the distance dummies and the DID variable in Fig. 1 (a). Compared to the price of residential units within 10 km of city centers, the price of units 10 to 20 km away from city centers increased by one percent after the COVID-19 outbreak. This effect increased to between 1.5 and 2% for units further from city centers. Fig. 1(a) clearly suggests the COVID-19 outbreak increased household willingness to pay for homes far away from city centers.

Fig. 1

The effect of COVID-19 on housing price gradients.

The effect of COVID-19 on the housing price gradient with respect to building height

To estimate the impact of COVID-19 on the housing price gradient measured by building height, we still follow Eq. (2). As displayed in column (1) of Table 3 , the coefficient of the building height variable implies that taller buildings are associated with lower housing prices. The coefficient of the triple interaction term is significantly negative, suggesting that apartments in taller buildings suffer from greater drops in housing prices compared to those in lower buildings. The corresponding event-study results in Online Appendix Fig. A5 indicate that household preferences for apartments in lower buildings gradually increased after the COVID-19 outbreak. This change is not driven by pre-trends and persists to the end of the sample period.

Table 3

The impact of COVID-19 on the housing price gradient with respect to building height.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	All projects	Projects with confirmed Covid-19 cases	Projects without confirmed Covid-19 cases	Projects with hospitals	Projects without hospitals	Strict social-distancing policies	Less strict social-distancing policies
Year 2020 (Treat)	0.001	-0.008	-0.002	0.011	0.009	-0.024**	0.012
	(0.006)	(0.020)	(0.007)	(0.009)	(0.009)	(0.011)	(0.008)
After SF (SF_t)	0.032***	0.008	0.035***	0.032***	0.036***	0.005	0.045***
	(0.005)	(0.016)	(0.006)	(0.007)	(0.007)	(0.009)	(0.006)
Treat×SF_t	0.004	-0.001	0.008	-0.001	0.008	0.032***	-0.009
	(0.006)	(0.018)	(0.007)	(0.009)	(0.009)	(0.011)	(0.007)
ln(height)	-0.056***	-0.045***	-0.058***	-0.042***	-0.062***	-0.058***	-0.056***
	(0.002)	(0.005)	(0.002)	(0.003)	(0.003)	(0.004)	(0.003)
Treat×ln(height)	0.017***	0.018***	0.017***	0.009***	0.021***	0.021***	0.017***
	(0.002)	(0.005)	(0.002)	(0.003)	(0.002)	(0.003)	(0.002)
SF_t×ln(height)	0.007***	0.012***	0.006***	0.004	0.009***	0.011***	0.005***
	(0.001)	(0.004)	(0.002)	(0.002)	(0.002)	(0.003)	(0.002)
Treat×SF_t×ln(height)	-0.009***	-0.011**	-0.009***	-0.009***	-0.010***	-0.016***	-0.006***
	(0.002)	(0.005)	(0.002)	(0.003)	(0.002)	(0.003)	(0.002)
Observations	747117	111521	614623	265390	481727	266745	480372
R-squared	0.979	0.980	0.979	0.979	0.979	0.951	0.984
Project FE	YES	YES	YES	YES	YES	YES	YES
Week pair FE	YES	YES	YES	YES	YES	YES	YES

Notes: All regressions follow Eq. (2). Housing attributes, the distance to city centers and its interaction terms with the DID variables are controlled but not displayed. represents whether the transaction date is after the Spring Festival in 2019 or 2020. denotes whether a transaction is in year 2020. The results in column (1) use all observations in the sample. In column (2), we define a project as having confirmed cases when there are confirmed Covid-19 cases within five hundred meters of the project centroid by April 8, 2020, when Wuhan reopened the city. Otherwise, projects are defined as not having confirmed cases in column (3). The results in columns (2) and (3) exclude four cities (Huhhot, Langfang, Shaoxing, and Shijiazhuang), because our dataset does not report the geocodes of confirmed Covid-19 cases in these four cities. A project is considered close to hospitals and included in column (4) if there is at least one hospital within 500 meters of the project centroid. Otherwise, a project is defined as far from hospitals and included in column (5). Column (6) includes cities that adopted stringent outdoor activity restrictions where residents were confined within projects. The other cities that did not adopt such policies are displayed in column (7). Standard errors in parenthesis are clustered at the project level. * indicates significance at the 10% level, ** indicates significance at the 5% level, and *** indicates significance at the 1% level.

Potential heterogenous effects are investigated in columns (2)–(7) of Table 3. The effect of COVID-19 on the housing price gradient with respect to building height is homogeneous for homes with and without confirmed COVID-19 cases and for those close to and far away from hospitals (see Online Appendix Table A5). Thus, household preferences for units in taller buildings decreased during the pandemic, regardless of the proximity of confirmed COVID-19 cases or hospitals.

The results in columns (6)–(7), however, indicate that the negative effect of COVID-19 on housing prices for units in taller residential buildings was significantly greater in cities with strict social-distancing policies. As mentioned earlier, the social-distancing policies in many cities restricted activities outside of a household's residential project. Thus, households in cities with strict social-distancing policies tended to substitute outdoor activities outside of projects with activities within projects. Home buyers in these cities were willing to pay a larger premium to live in projects with lower residential buildings in order to reduce the infection risk.

Similar to the analysis in Section 4.2, we discretize the building height variable and report the estimates for the interaction terms between the building height dummies and the DID variable in Fig. 1(b). We categorize building heights into five groups, according to the distribution of the total number of floors shown in Online Appendix Fig. A4. In general, we observe decreasing housing prices with respect to building height after the COVID-19 outbreak. The prices of apartments in buildings with more than 30 floors decreased by about two percent compared to those of apartments in buildings with fewer than six floors. This is consistent with our previous findings that households paid less for apartments in buildings with higher density in order to avoid the risk of infection.

We further explore the heterogenous effect on the availability of elevators and the density of apartments on each floor. Using elevators in a crowded and enclosed indoor space is likely to increase the infection risk. Columns (1) and (2) in Online Appendix Table A6 show that the change in household preference for vertical density measured by building height is driven by residential buildings with elevators. These results are also consistent with our findings in Online Appendix Fig. A4, because buildings with elevators are usually taller buildings. We then use the ratio of the number of housing units on each floor to the number of elevators in the building to measure the density on each floor. More apartments on each floor is correlated with a higher infection risk of COVID-19, as households must share stairs or elevators. Column (6) suggests that COVID-19’s effect on the housing price gradient with respect to building height is greater for buildings with elevators that have more housing units on each floor. These results serve as further support for the previous finding that COVID-19 decreased household preferences for denser residences.

The effect of COVID-19 on the housing price gradient within residential buildings

The empirical specification still follows Eq. (2), but the variable is now replaced by two dummies representing the range of the unit floor level. The “low floor” category is omitted as the base group. In this specification, we report estimates on whether the prices of high-floor and middle-floor units changed after the COVID-19 outbreak compared to those of low-floor units. Distance to the city center and its interactions with the DID estimator are still included in the regressions. The regression results are summarized in Table 4 .

Table 4

The impact of COVID-19 on the housing price gradient within residential buildings.

	(1)	(2)	(3)		(4)	(5)	(6)
	Buildings with different heights (measured by the total number of floors)
	<=11 without elevators	<=11 with elevators		<=11	12-18	19-30	>=31
Year 2020 (Treat)	0.055***	0.075***		0.057***	0.049***	0.053***	0.028***
	(0.007)	(0.018)		(0.006)	(0.010)	(0.009)	(0.009)
After SF (SF_t)	0.054***	0.070***		0.056***	0.054***	0.064***	0.046***
	(0.006)	(0.017)		(0.006)	(0.008)	(0.008)	(0.007)
Treat×SF_t	-0.023***	-0.027		-0.025***	-0.015	-0.037***	-0.021***
	(0.007)	(0.018)		(0.006)	(0.010)	(0.009)	(0.008)
Middlefloors	-0.011***	0.019**		-0.005	0.033***	0.030***	0.023***
	(0.004)	(0.009)		(0.004)	(0.005)	(0.005)	(0.005)
Treat×Middlefloors	0.002	-0.008		-0.001	-0.011*	-0.013**	-0.005
	(0.005)	(0.010)		(0.005)	(0.006)	(0.006)	(0.006)
SF_t×Middlefloors	-0.004	-0.009		-0.005	-0.007	-0.011**	-0.006
	(0.004)	(0.009)		(0.004)	(0.005)	(0.005)	(0.005)
Treat×SF_t×Middlefloors	-0.004	0.005		-0.001	0.008	0.011*	0.005
	(0.005)	(0.011)		(0.005)	(0.006)	(0.006)	(0.006)
Highfloors	-0.077***	0.017*		-0.059***	0.025***	0.025***	0.022***
	(0.004)	(0.010)		(0.004)	(0.006)	(0.005)	(0.005)
Treat×Highfloors	-0.015***	-0.020*		-0.014***	-0.009	-0.007	-0.005
	(0.005)	(0.012)		(0.005)	(0.007)	(0.006)	(0.006)
SF_t×Highfloors	-0.019***	-0.022**		-0.017***	-0.004	-0.009	-0.010*
	(0.004)	(0.010)		(0.004)	(0.006)	(0.005)	(0.005)
Treat×SF_t×Highfloors	0.013**	0.022*		0.013***	0.009	0.009	0.007
	(0.006)	(0.012)		(0.005)	(0.007)	(0.006)	(0.006)
Observations	225475	72484		303488	131849	156047	147290
R-squared	0.981	0.977		0.980	0.984	0.982	0.974
Project FE	YES	YES		YES	YES	YES	YES
Week pair FE	YES	YES		YES	YES	YES	YES

Notes: All regressions follow Eq. (2). Housing attributes, the distance to city centers, and its interaction terms with the DID variables are controlled but not displayed. represents whether the transaction date is after the Spring Festival in 2019 or 2020. denotes whether a transaction is in year 2020. Column (3) includes a control variable indicating whether the house has access to elevators. Standard errors in parenthesis are clustered at the project level. * indicates significance at the 10% level, ** indicates significance at the 5% level, and *** indicates significance at the 1% level.

As the exact floor level of two units in the same floor level range can be quite different depending on the building height, we run separate regressions for buildings with different heights. For buildings with less than 11 floors, we further split the sample according to the availability of elevators, as some low buildings do not have them, and the vertical pattern within low-rise buildings may vary by having access to elevators or not. All taller buildings are equipped with elevators.

Column (1) displays the results for apartments in low-rise buildings without elevators. Among these apartments, the price of those on middle and high floors is 1.1 and 7.7 percentage lower than that of units on low floors, respectively. This is consistent with the intuition that housing prices decrease with floor level in buildings without elevators in which one can only reach higher units by climbing stairs. In column (2), we find that apartments on “middle floors” are the most expensive in buildings with elevators, followed by units on “high floors” and finally those on “low floors”. We then pool properties in low-rise buildings in column (3) and control for the availability of elevators. Results are similar to those in column (1) as most low buildings do not have elevators. Across columns (4) to (6), we find similar patterns for homes on different floors within taller buildings. However, compared to column (2), the price premium of properties on “middle floors” and “high floors” are more significant and greater. Given that buildings have access to elevators, this suggests that the vertical gradient within a residential building is more apparent in very tall buildings.

The coefficients on the triple interaction terms in column (1) suggest that COVID-19 significantly increased the prices for units on “high floors” relative to the prices of those on “low floors” in buildings without elevators. The triple interaction terms in column (2) suggest a similar finding in low-rise buildings with elevators, but this estimate is only statistically significant at the 10% level. The results in column (3) confirm that the relative prices of apartments on high floors increased in all low-rise buildings. For apartments in buildings with a total number of floors greater than 12, columns (4)–(6) suggest that COVID-19 also increased the relative price of units on middle and high floors, although these estimates are statistically insignificant. However, the estimates of the triple interaction terms in columns (4)–(6) are not statistically different from the corresponding estimates in columns (1)–(3), thus indicating that COVID-19 changed the vertical gradient within residential buildings in a similar way, regardless of building height.

Conclusions

This paper finds that the COVID-19 pandemic has had a salient effect on the Chinese urban housing market. The average housing price in 60 cities decreased by about two percent immediately after the COVID-19 outbreak but gradually recovered by September 2020. More importantly, the estimates in this paper suggest that COVID-19 reduced household preferences for both horizontal and vertical urban density.

The identified effects in this paper should not be explained as long-term impacts. Due to lack of data, we are not able to study the performance of the Chinese housing market over a longer period. However, it is intriguing to observe that COVID-19 has affected the public's preferences for urban density even within this short period of time. The event-study results suggest household aversion for urban density persisted during our sample period. How long will this change in preferences last? Will COVID-19 permanently change the pattern and scope of agglomeration economies? These questions are beyond the scope of this research and may be answered in future studies.

CRediT authorship contribution statement

Naqun Huang: Conceptualization, Methodology, Software, Validation, Formal analysis, Data curation, Writing – original draft, Writing – review & editing. Jindong Pang: Conceptualization, Methodology, Software, Validation, Formal analysis, Data curation, Writing – original draft, Writing – review & editing. Yanmin Yang: Conceptualization, Methodology, Software, Validation, Formal analysis, Data curation, Writing – original draft, Writing – review & editing.