How working from home reshapes cities

Significance

The future of cities is a topic of broad interest to businesspeople making decisions around office space, policymakers planning zoning codes and public transportation networks, and academics interested in questions of spatial organization across a wide range of fields including economics, urban planning, environmental studies, and sociology. Our data provide evidence as to how a huge structural shift to work—remote working—may have permanently changed the shape of many major global cities and is therefore of value to these groups.

Abstract

In recent decades, economic activity has become increasingly concentrated in major global metropolises. Yet, the rise of working from home threatens this dominance of cities. Using multiple high-frequency datasets on spending, commuting, migration, and housing, we provide global evidence that remote work has dispersed economic activity away from city centers. We label this the “Donut Effect,” which is much larger and more persistent in cities with high levels of remote work. Using detailed household microdata from the United States, we show that three-fifths of households that left city centers in big cities moved to the suburbs of the same city. This is likely explained by the rise of hybrid work, in which employees still commute to the office a few days a week. The enduring popularity of hybrid work into 2024 suggests that the Donut Effect will persist while also leaving broader metropolitan areas intact.

How will humans organize themselves across space in the future? In 1800, less than 10% of humans lived in urban areas. Today around 4.4bn people live in cities—over half the world’s population (1). Technologies from factories to ports played a central role in the formation of cities by increasing the economic returns to density. Yet, technology can also make it easier for humans to live further apart. The automobile led to the suburbanization of America in the mid-20th century. More recently, information technology has made it easier to collaborate at a distance, leading to repeated predictions that it could one day reduce the need for cities. For instance, in 1995, the technology writer George Gilder claimed that the internet revolution would lead to the “death of cities” (2). In fact, the next two decades marked the remarkable growth of “superstar” cities from New York City to Beijing (3, 4). When the Covid-19 pandemic struck in 2020 and cities quickly hollowed out, analysts again predicted their demise (5).

Four years later, city planners, researchers, and business leaders are wondering whether urban centers are on track to return to their prepandemic states or if we are in a new normal (6–12). Such questions are motivated by an ongoing debate over the future of work-from-home (WFH). In the United States, the share of days worked from home rose from 5% prepandemic to 60% in the summer of 2020, fell throughout 2021 and 2022, but then stabilized at nearly 30% of days in 2023 and 2024 (SI Appendix, Fig. S1) (12). Office occupancy rates and foot traffic to commercial offices across major US cities remain depressed at around 30 to 50% below prepandemic levels, with some data sources showing stabilization and others showing a gradual, albeit slowing, recovery (13, 14).

In theory, WFH has large effects on city structure (15–20). By reducing commutes, WFH enables an employee to live away from their place of work. How exactly this plays out depends on the nature of WFH. If it is a few days a week (hybrid), then workers are still weakly tethered to their place of work. If WFH is full-time, then workers are free to move where they choose, subject to some constraints around time zones and immigration laws. Either way, WFH should impact commercial city centers, areas like Manhattan in New York City that feature a high concentration of economic activity (21, 22). In this paper, we provide evidence that cities around the world have experienced a “Donut Effect” in which economic activity disperses away from city centers. We argue this is driven by the widespread adoption of hybrid WFH.

Thus far, the empirical literature has focused on the short-term effects of the pandemic and the associated surge in WFH on cities (23–28). Due to data limitations, most studies focus on the United States, though a few have examined other countries in isolation including Spain and the United Kingdom (29–31). Ref. 32 examines housing markets across 16 OECD countries and finds evidence of a shift in housing demand towards suburbs in data till 2021. But to our knowledge, there has been no systematic evaluation of the assocation between WFH and shifts in urban structure across a large sample of countries. Little attention been paid to the heterogeneity of effects across different types of cities—why some cities experience strong effects whereas others experience little if anything. Finally, existing work has not distinguished between the effects of different types of WFH (hybrid vs full-time). Our study fills these gaps.

Our main dataset in this paper covers aggregated in-person transactions on the Mastercard network across 118 global cities, which we use to measure consumer spending patterns. We supplement this with a variety of US-specific datasets to enrich our results on the heterogeneity and permanence of the Donut Effect. These include migration patterns (USPS, Data Axle), US transportation patterns (Inrix, National Transportation Database), and US real estate markets (Zillow). Migration flow data from the USPS are derived from change-of-address records and are publicly available but are largely absent from academic research because they were only released in 2020. Three datasets (Mastercard, Data Axle, Inrix) are proprietary and until this paper have not been used in our setting.

Our study is motivated by several key questions: i) What is the impact of WFH on economic activity in city centers around the world in the short and long-run? ii) How do these changes vary within and across different types of cities? iii) Finally, what do the changes to city structure imply for the future of WFH–will it be hybrid or fully remote?

Materials and Methods

Here, we provide a brief description of our key datasets and approaches to measuring the Donut Effect, with more details in SI Appendix, Supplementary Materials. Data and replication code is available at ref. 33.

Consumption Spending—Mastercard.

We use the aggregate in-person transaction on the Mastercard card network to track consumption spending in cities from January 2018 to September 2023. For privacy protection reasons, we report aggregated spending indices within defined geographic areas. We construct a monthly “donut index” by taking the level of spending in the city center, defined as a 2-mile radius circle around the city center’s coordinates, and subtracting the spending for the remainder of the city, defined as the ring from 2—N miles out where N varies from 5 to 50 in increments of 5. For our main figures we fix N = 30, but our results are robust to all other values. We normalize each index such that the average 2019 value = 100 before differencing. The resulting index measures the relative growth rate in spending between the city center and the surrounding region.

Migration Flows—US Postal Service (USPS) and Data Axle.

USPS’s National Change of Address dataset contains monthly data on population flows at the zip-code level from April 2017 to June 2023. The data capture when a US household moves address. In our main figures, we construct time series of net population inflows as share of prepandemic population for four groups of zip codes within each city: the city center, high density, the suburb, and the outer suburb. Our boundaries for each city are given by US Census Bureau Metropolitan Statistical Areas (MSAs).

USPS does not link the origin and destination location of moves. We use quarterly files containing the near universe of US households from 2018 to 2021 from Data Axle, a data services firm. Each file contains each household’s latest address and other demographic and income characteristics. We link households over time to measure the number of households moving between pairs of zip codes.

Real Estate Markets—Zillow.

Shifts in the demand for location should show up in real estate markets. We measure this using Zillow’s home value index (ZHVI) and observed rental index (ZORI). We take population-weighted averages of the indices and bucket them into four groups: the city center and the three population density buckets. Our main figures normalize the series to 100 in Feb 2020. See (34) for more details on Zillow’s indices and our SI Appendix, Supplementary Materials for robustness checks.

Commuting—Inrix.

WFH reduces the need to visit the city center for work regardless of whether people move. To test this, we use point-to-point trip data for all GPS-connected cars from one of America’s largest car manufacturers. Our data cover all trips from January to June 2019 and January to June in 2023 inclusive. We define a commute as any trip that starts outside the city center (defined as a 2-mile radius around the city center coordinates) and ends within the city center. In our main figures, we plot the frequency distribution of commutes over the hours of an average weekday or weekend day, as well as the average speed of commuting.

Working-from-home (WFH)—Various Sources.

We use three datasets to measure WFH in cities across the world and across cities in the United States. Globally, we rely on survey data from the Global Survey of Working Arrangements (G-SWA) which contains WFH shares for 34 countries (35). In the United States, our first measure looks at prepandemic WFH exposure by merging the distribution of workers across industries from the US Census (LODES) with assessments of the share of jobs that can be done from home in each two-digit NAICS industry from (36). Our second measure for the United States looks at actual WFH rates using the near universe of jobs-postings from the WFH Map program (37).

Results

Fig. 1A highlights the Donut Effect in consumer spending for 118 major cities across 50 countries. It plots the difference in spending growth between the city center and the suburbs 5, 10, 30, and 50 miles out. This gap shows a sharp drop in March 2020, which attenuates in 2021 and 2022, and then stabilizes in 2023 at about −15 percentage-points (pp). So, on average, suburbs in 118 large global cities have seen spending grow 15 pp more relative to their city centers, affecting hundreds of billions of dollars of global consumer spending (New York City alone had $101bn in taxable retail and consumer services sales in 2022).

Fig. 1.

Global cities see a shift in consumer spending shares from city centers to suburbs. Panel (A) shows the donut index constructed using Mastercard spending data for 118 cities across 50 countries (max 3 cities per country). We receive spending indices for concentric circles that represent the spending index of the city center (2 mile radius circle) subtracted from the spending index of the outer ring (2 miles to N miles). Each spending index is normalized such that the average 2019 value is 100; thus, the difference has an average value of 0 in 2019. The level of the index can be interpreted as the relative growth of the city center vs the outer ring. We plot a population-weighted average index across cities. Panel (B) shows the shift in the distribution of consumer spending across zip codes in the New York metro area in percentage points. We take the distribution in the first 6 mo of 2023 and subtract the spending distribution of the first six months of 2019. Sources: Mastercard. Data: Jan 2018–Sep 2023.

As an example, we display a heat map for New York City in Fig. 1B. The map is colored by the percentage-point difference in a zip code’s share of overall city spending from the first half of 2019 to 2023. Most zip codes in Manhattan are red, indicating a reduction in spending share of up to 0.5 pp. In contrast, the suburbs are mostly green in color reflecting the substantial increases in spending, with increases up to 0.5 pp in the suburbs of New Jersey, Brooklyn, Long Island, and Queens. Other major global cities show similar patterns of drops in city center spending and rises in their surrounding suburbs, both in developed and emerging economies (SI Appendix, Fig. S2). There is no obvious regional or continental pattern to these results, though it is notable that Western Europe cities have outperformed the rest of the world (SI Appendix, Fig. S10B).

One interpretation of these data is that WFH reduces the frequency of office-going, which naturally reduces consumer spending on amenities in city centers. Commutes to city centers have fallen, adding support to this story (SI Appendix, Fig. S7). Restaurant spending appears to have fallen more than spending on groceries and apparel, which plausibly reflects less office-going (SI Appendix, Fig. S10A). US-specific data suggest that there are deeper changes to city structure—namely a large reallocation of people and businesses out of city centers, as shown by data on population flows, business flows, and real estate prices. We do not have similarly comprehensive population and housing price data for cities in other countries, making this a ripe area for future work. However, there are many global cities that have Donut Effects of comparable magnitudes to that of major US cities, suggesting that they experienced both effects. For example, Sydney, Toronto, and Berlin all had bigger effects than even New York City (SI Appendix, Fig. S2).

The movement of people out of city centers amplifies the WFH-induced Donut Effect. Fig. 2A plots cumulative net inflows of people as a share of 2019 population after adjusting for prepandemic trends. City centers in the top 12 US cities see cumulative net outflows of 8% of their 2019 population. This reflects an additional 170,000 people out of the initial 2 m city center inhabitants in 2019. The heat map for New York City in Fig. 2B visually shows how the pattern of movement is shaped like a Donut—with outflows from the city center and inflows to surrounding areas.

Fig. 2.

Big US cities have experienced large population outflows from city centers. Panel (A) shows monthly net population inflows summed across all zip codes in a bucket before dividing by total 2019 population from the 2015–19 ACS. We then take the difference from the average flow in two years prepandemic (2018 and 2019) before cumulating starting in Jan 2018. Zip codes are grouped by population density bucket or presence in the city center as follows: high density = top 10%, suburb = 50 to 90th percentile, exurb = 0 to 50th percentile. The city center is defined by taking all zip codes with centroids contained within a 2 mile radius of Central Business District coordinates taken from Holian (2019). Panel (B) shows net population inflows cumulated from Feb 2020 to Jun 2023 as a share of 2019 population. Sources: USPS, Census Bureau, Holian (2019). Data: Jan 2018–Jun 2023.

Where have these city-center leavers gone? A common view is that the urban exodus has led to a boom of fully remote workers in rural areas (38). Indeed, small cities and rural areas have seen population growth totaling 1 to 2% of their prepandemic populations since 2020 after adjusting for prepandemic trends (SI Appendix, Fig. S6). However, this is not the dominant pattern of movement for households leaving big cities. Using quarterly panel data on US household addresses from Data Axle, and after adjusting for the prepandemic pattern, we find that 58% of households that left the centers of big cities just moved further out within the same city. Breaking down this group, 22% went out to high-density areas, 13% moved to mid-density suburbs, and 23% to low-density suburbs. The remaining 42% who left city centers mostly relocated to other large top-12 sized cities (9%) or smaller cities sized 13 to 340 (29%). Just 4% move to rural areas (SI Appendix, Fig. S5). One reason for this pattern is that most remote work is carried out by hybrid employees, coming in typically two or three days a week rather than fully remote employees (12, 35). Hybrid employees need to live within a reasonable commuting distance from their offices. In our SI Appendix, Supplementary Materials, we account for shifts in the inflow of households into city centers to calculate net outflows. We find more sizable shifts to other metro areas, though these are still small as share of destination population.

Interestingly, the Donut Effect appears to be limited to large cities. Fig. 3 examines population flows in the largest 12 US cities, cities ranked 13 to 50, and cities below the top 50 by population. The largest cities, for example Atlanta, Chicago, New York City, and San Francisco, saw large city center outflows that have not reversed. The medium-sized cities, places like Cleveland, Indianapolis, and Nashville saw only minor movement out from their city centers. The smaller cities, places like Des Moines, Boulder, and Salem have not seen any measurable exodus. One reason for these differences is that larger cities have greater concentrations of business service, finance and tech graduate jobs which have high levels of WFH due to the computerized nature of the job (22, 39). In contrast, smaller cities are more manufacturing, retail, and wholesale focused which are less amenable to remote work. Smaller cities also have flatter city center to suburb price gradients, reducing the incentive for residents to move out to the suburbs to purchase larger properties. These results are consistent with ref. 26 which finds that across the top 30 US metro areas, WFH exposure is correlated with the degree to which the city center-suburb home price gradient flattens after the pandemic.

Fig. 3.

Big US cities see large outflows from city centers; small cities see none. All three panels shows monthly net population inflows divided by 2019 population from the 2015–19 5-yr ACS. Panel (A) pools the top 12 metros by population, panel (B) contains metros 13 to 50. and panel (C) gives the remaining metros (we have data on 340 in total). We then take the difference from the average flow in two years prepandemic (2018 and 2019) before cumulating starting in Jan 2018. Zip codes are grouped by population density or presence in the city center as follow: high density = top 10%, suburb = 50 to 90th percentile, exurb = 0 to 50th percentile. The city center is defined by taking all zip codes with centroids contained within a 2 mile radius of Central Business District coordinates taken from Holian (2019). Sources: USPS, Census Bureau, Holian (2019). Data: Jan 2018–Jun 2023.

Impact on Real Estate and Commuting.

WFH benefits real estate markets further from city centers in two ways. First, by reducing commute frequency, it makes it more convenient to live further away from one’s place of work. Our data on GPS-connected trips from a major US auto manufacturer show a 20% reduction in vehicle trips from the suburbs to the city center during rush hour (7 to 8am). Driving speeds rose by about 10% during this period with the drop in traffic volumes. Total US driving miles dropped by around 6% while rail transit journeys fell by a greater 30% due to its higher reliance on commuting. These changes have also persisted till the end of 2023 (SI Appendix, Figs. S7 and S8). Second, WFH increases the premium to space, leading to greater demand for larger suburban homes (40).

Zillow data show the impact on real estate markets. Fig. 4A shows a 15-pp gap in housing rental growth since March 2020 for the twelve largest US cities. This gap stabilized in 2023 as levels of WFH stabilized. Zillow’s home value index, which estimates home values, shows an even larger divergence between the city center and suburbs (Fig. 4B). While city center prices were falling slightly prepandemic, the pandemic marks the start of a striking divergence between high-density regions and the suburbs that reaches over 40 percentage points by September 2023. Housing prices, like all asset prices, are forward-looking, incorporating expectations of future demand (26, 27). So, this large growing gap in housing prices between the city center and suburbs suggests that real-estate markets expect an enduring increase in the demand for suburban housing over city centers.

Fig. 4.

Real estate demand has shifted from US city centers to suburbs. Panel (A) shows Zillow’s observed rental index and panel (B) shows the home value index in the 12 largest US metro areas (New York, Los Angeles, Chicago, Dallas, Houston, Miami, Philadelphia, Washington DC, Atlanta, Boston, San Francisco, and Phoenix—ordered by population). Zip codes are grouped by population density or presence in a Central Business District (CBD). A population-weighted average is taken across all zip codes in each bucket, and each aggregated index is normalized such that Feb 2020 = 100. Groups are given by high density = top 10%, suburb = 50 to 90th percentile, exurb = 0 to 50 percentile. The city center is defined by taking all zip codes with centroids contained within a 2 mile radius of Central Business District coordinates taken from Holian (2019). Population data taken from the 2015–19 5-yr ACS. Sources: Zillow, Census Bureau, Holian (2019). Data: Jan 2018–Dec 2023.

Explaining the Donut Effect.

The rise of remote work is a key factor explaining the magnitude and persistence of the Donut Effect. Prepandemic, the share of days worked from home was around 5% in the United States. After spiking during the height of the pandemic, the share has stabilized around 27% in survey data till June 2024 (12). Fig. 5 panel A bins 28 global cities into two groups based on the surveyed rate of WFH in each city’s country (35). Our measure of the Donut Effect for each group is calculated as a population-weighted average of the city-level difference in normalized spending between the city center and the outskirts. The high-WFH group (top third of cities) has a large and persistent effect that has flatlined from early 2022 till the end of 2023 at around 20 pp.

Fig. 5.

The shift in spending shares out of city centers is larger for high-WFH cities. Panel (A) buckets Mastercard spending data from 25 global cities into two groups based on observed WFH shares taken from ref. 35 (the top third of cities and the remainder). For each city, we take the city center spending index (2 mile radius around the city center) and subtract a spending index for the ring from 2 to 30 miles. We normalize the average 2019 value to 100 before subtracting. A population-weighted average is taken across cities in each bucket. Panel (B) buckets 144 US metro areas into two groups based on observed WFH shares taken from ref. 37 (top third and the remainder). Net population inflows as a share of 2019 population are differenced from the average net inflow in the two years prepandemic (2018 and 2019) before cumulating starting in Jan 2018. The city center series is differenced from the rest of the metro before plotting. Sources: Mastercard, USPS. Data: Jan 2018–Sep 2023.

Panel B shows the relationship between WFH and the Donut Effect in consumer spending across US cities. WFH rates are taken from up-to-date survey data from ref. 37 and combined with data on the cumulative population outflows from city centers relative to the outskirts (area outside the city center). We bucket 144 metros into a high-WFH group with the top third of metros and a low-WFH group with the bottom two-thirds. The high-WFH group sees population outflows that are 5 pp greater for city centers than outskirts after adjusting for prepandemic trends. This measure has stabilized since 2022, whereas the low-WFH group only sees a 0.5 pp difference in outflows.

To validate these findings, we test the global cross-city relationship between WFH and Donut size across 28 global cities. There is a positive and significant relationship (t-stat 2.3; P-value 0.03), but statistical power is limited because of a small sample size (SI Appendix, Fig. S9A). US data bolster the result. Across the top US 100 cities, there is a stronger relationship between the size of the Donut Effect and observed levels of WFH (t-stat 3.1; P-value 0.00) (SI Appendix, Fig. S9B). We also run US and global cross-city regressions with density, cumulative lockdowns, and GDP per capita as controls (SI Appendix, Table S11). WFH has the expected negative sign in both regressions but only remains significant in the global regression.

Next, we look at the factors explaining the Donut Effect within the United States at a more granular level. We run zip-code level regressions of changes in rents, home values, populations, and business against the share of population that can WFH with metro-level fixed effects. The regressions show statistically significant negative relationships between WFH and changes in rents, home prices, population flows, and business flows (SI Appendix, Tables S12 and S13). In particular, a 1% increase in the share of residents that can WFH in a zip code is associated with .15 point reduction in the rental index, a .13-point reduction in the home value index, a .07 percentage-point reduction in population inflow, and a .11 percentage-point decrease in business establishment inflow. Adding controls for density and distance to the city center makes the WFH coefficient in the home value and business flow regressions insignificant since the two variables are correlated with WFH.

Finally, we examine how the magnitude of the Donut Effect in population flows changes further out from the city center (SI Appendix, Fig. S4). We find that it grows up to 10 to 15 miles away from the city center before attenuating. Our explanation is that households move out from the city center to access lower real estate costs (24–26), but eventually, commute costs become prohibitive because they must still go to work a few days a week.

Our broader interpretation of these results is that WFH mediates the long-term impact of lockdowns on cities, shifting some cities into a new equilibrium (20). Identifying the precise causal effect of WFH is challenging because the WFH shock is coincident with pandemic lockdowns that may have made shifted urban structure for other reasons. For example, if cities have two equilibria, low-WFH and high-WFH, and the pandemic shifted cities from one to the other, one could describe this as a change in urban structure leading to a change in WFH, rather than the other way around. Nonetheless, a shift in urban structure is not possible unless a city has high WFH-potential, as defined by the share of jobs in a city that in principle can be done from home. In our data, observed WFH is highly correlated with measures of WFH-potential (36). In addition, cities with high rates of WFH have had larger and more persistent Donut Effects (Fig. 5).

This correlation is a result of two mechanisms. First, WFH directly reduces commute frequency leading to less spending in cities. Second by reducing commute costs, WFH leads to increased movement out of city centers, which further reduces spending in cities. A precise decomposition of the reduction in spending between movers and nonmovers is a subject for future work. However, our data, which show large spending reductions, reduced commute frequency, sizable population outflows, and shifts in real estate demand, suggest both effects are important.

Persistence.

A quick analysis of our figures suggests that rates of remote work and measures of the average Donut Effect in spending have stabilized in large cities with high rates of WFH. In particular, the drop in average city center spending across our 118 global cities stabilized in late 2022 (Fig. 1). Movements in population across the top 12 US cities have not reversed as of mid-2023 (Fig. 2). Rent and house price gaps across the top 12 US cities remain stable till the end of 2023 (Fig. 4). In addition, levels of working from home have been stabilized as of April 2024 (SI Appendix, Fig. S1). As Fig. 5 shows, in low and mid-WFH cities, these changes have largely reverted, but in high-WFH cities, they have remained large and persistent. This is true in the global Mastercard data and the US-specific USPS data.

Extrapolating from the recent past to the future depends on how WFH evolves. There is considerable debate over this. Some data suggest that WFH will gradually subside. Data from Placer.ai, a location analytics firm, which tracks foot traffic to 1,000 commercial office buildings in major US cities show that visits were down 32% in May 2024 compared to May 2019 (14). In May 2023, the same figure stood at 38% and in May 2022 at 50%. So while foot traffic is growing, the pace of that growth is slowing. Some companies have also required full-time office attendance for parts of their workforce including JPMorgan Chase, UPS, and Boeing (41). In theory, there may also be long-term costs to WFH that are currently hard to detect—for example, training of new employees may suffer and innovation may prove more difficult with less in-office time.

Yet, other evidence suggests that WFH has stabilized and may even grow in the longer-run. Survey data on WFH rates suggest that they have stabilized at just under 30% of days since mid-2023 (12). Scoop’s data finds that 38% of firms in their sample required full-time office presence by the end of 2023, down from 49% at the start of the year (42). A randomized trial of hybrid-WFH (43) shows that hybrid work leads to more employee satisfaction and retention without sacrificing productivity, so it could prove to be a competitive advantage for firms. Patent applications for technologies that support WFH have grown (44), reflecting the induced response to a larger market for software that benefits remote work. Firms are iterating on WFH-related management practices, so the organizational cost of WFH may fall over time. The net effect of these factors is an empirical question, but our assessment of the data thus far is that at the very least a considerable portion of the rise in hybrid-WFH will persist.

Discussion

WFH had notable effects on the structure of global cities during the pandemic. Yet the long-term effects of those changes are still debated, with many arguing cities are reverting to their prepandemic states. Using six datasets across real estate markets, migration flows, commuting patterns, public transportation usage, and consumption spending, this paper argues that our core finding of the Donut Effect—a reallocation of economic activity from city centers to suburbs for highly agglomerated cities—appears to have stabilized.

We characterize the Donut Effect in several ways. We show that it is larger for the biggest US cities, and almost completely absent in the smallest. We also show that WFH is correlated with the size of the Donut both across cities within the United States and across major cities worldwide. Our household address data show that most households leaving city centers go to the suburbs of the US same city rather than another city or rural area. In work that this paper builds upon (45), we built a simple spatial equilibrium model which analytically confirms this is consistent with the now dominant hybrid WFH pattern of knowledge workers. Our interpretation is that WFH weakens the type of agglomeration economies that lead to concentrations of economic activity in city centers but preserves the type of agglomeration economies that lead to clustering in broader metro areas (3, 4, 39).

This paper raises several policy implications. The Donut Effect may cause the tax base to shift from cities to surrounding areas, putting strain on large cities’ public finances. The decline in public transportation usage may put further strain on public finances and raises concerns about the longer-run solvency of mass transit systems. Finally, the drop in demand for offices is causing large reductions in office valuations (46), placing strain on real-estate firms and the banks that have financed them. On the positive side, many cities are actively working to rezone office space into mixed use space to attract residents back to city centers, which could make cities more livable. Commute times and peak-hour traffic have fallen, which are also positive for welfare. Finally, many employees have moved to less dense suburbs with more spacious homes and suburban amenities. Altogether, the net welfare effects of the Donut Effect are still ambiguous and will depend on how society continues to respond.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Data and code for this paper are included in an online github repository at https://github.com/arjunramani3/donut-effect-pnas (33) with the following exceptions: 1) Data Axle: contains the universe of US households with addresses from 2017Q1 to 2021Q4. This data is only accessible through a secure server hosted at the Stanford Graduate School of Business. We only report aggregated summary stats as part of a data agreement with Data Axle. The full data processing script is included in our replication file. 2) Inrix commuting data: contains the universe of GPS-connected car trips from a major US auto manufacturer. We only report aggregate summary stats as part of an agreement with Inrix. The full data processing script is included in our replication file. 3) Mastercard data: contains the near universe of in-person transactions on the Mastercard card network. We only report aggregated spending indices to track how city centers perform relative to suburbs as part of a data agreement with Mastercard. The full data processing script is included in our replication file.