Durable Change in U.S. Urban Mobility Networks,2019–2022

Abstract

In 2020, the COVID-19 pandemic significantly altered how people move between neighborhoods. Tracking these changes is important because a growing literature demonstrates that mobility networks influence social and environmental exposures that interact directly with urban inequalities. Using four years of weekly smartphone-based mobility data in the 25 largest U.S. cities, we investigate how mobility changed in 2021 and 2022. We measure mobility networks with three previously used indices and introduce a fourth, the Dissimilar Mobility Index, to capture the demographic dissimilarity experienced in a mobility network. We find that although mobility hubs and their associated patterns of segregated mobility returned to pre-pandemic levels in 2021, neighborhood isolation remained depressed until the end of 2022 compared to 2019. Together, these results indicate that despite vaccine availability in 2021, structural changes in urban mobility networks caused by the COVID-19 pandemic were durable for over two years after its onset.

In 2020, the COVID-19 pandemic had many consequential societal impacts (Furceri et al. 2022; Johnson, Joyce, and Platt 2021). Among them, researchers documented a massive disruption in daily movement throughout American cities (Gao et al. 2020; Huang et al. 2022; Toger et al. 2021; You 2022). The disruption in mobility meant not only a decline in total movement as a result of lockdowns and stay-at-home orders but also structural changes to mobility networks (Marlow, Makovi, and Abrahao 2021). Where people move throughout the day defines the social patterning of exposures to everything from air pollution (Brazil 2022; Ma et al. 2020) through crime (Graif et al. 2021; Graif, Lungeanu, and Yetter 2017; Levy, Phillips, and Sampson 2020), and economic opportunity (Covington 2018; Ruef and Grigoryeva 2020; Sugie and Lens 2017). Furthermore, daily mobility is also closely linked to residential segregation (Candipan et al. 2021; Phillips et al. 2021; Vachuska 2023; Wang et al. 2018), and this intersection places those who are disadvantaged both in their residential context and movement at a compounded disadvantage (Krivo et al. 2013; Levy et al. 2020). Given the societal implications of changes in travel within cities, this article investigates how the shifts that emerged at the beginning of the pandemic in 2020 (Gyorgy et al. 2023; Marlow et al. 2021) evolved with pandemic conditions.

In the context of COVID-19, monitoring changes in mobility has been critical because movement throughout cities was closely linked to disease spread (Badr et al. 2020; Glaeser, Gorback, and Redding 2022; Nouvellet et al. 2021). More recent research also links unequal patterns of mobility and residential segregation to the emergence of unequal case rates among neighborhoods (Chang et al. 2021; Gyorgy et al. 2023; Levy et al. 2022). Levy et al. (2022) for example, find that neighborhood COVID-19 case counts in several cities were exacerbated by higher numbers of trips to economically disadvantaged neighborhoods. In fact, the disadvantages present in places that a neighborhood was connected to by daily trips were more predictive of case rates than a neighborhood’s own level of disadvantage. Therefore, it is important for public health and societal well-being to understand how patterns of daily mobility change over time.

This article makes several contributions to the study of urban mobility inequality by using weekly data on the number of trips between census tracts in the 25 largest U.S. cities. First, we employ three previously studied mobility indices to show how these measures evolved over four years (2019–2022). Second, to complement these existing indices, we introduce a new measure that we call the Dissimilar Mobility Index (DMI). The DMI captures segregated mobility by measuring the average level of sociodemographic dissimilarity present in a mobility network. Earlier work explored how mobility networks changed in the initial phases of the COVID-19 pandemic (Huang et al. 2020; Marlow et al. 2021; Toger et al. 2021; You 2022), but here we show for the first time how mobility networks continued to evolve in 2021 and 2022. In particular, mobility hubs, such as downtown business districts, returned to their prepandemic prominence as popular destinations in 2021 but declined again in 2022. Racially segregated mobility also appears to have returned to 2019 levels in 2021. Importantly, however, the measure of neighborhood isolation indicates that places remained more isolated from one another throughout 2021 and 2022 compared to 2019. Thus, the structural changes to mobility networks that emerged during the pandemic have been durable in at least one important way.

Data

We use two main sources of data. First, we use mobility data from the largest 25 U.S. cities to construct our index measures of mobility networks from 2019 to 2022. We supplement the mobility data with census data on sociodemographics. We identify cities by selecting counties fully containing the boundaries of a selected city. In cases where a city boundary extends to multiple counties—such as New York City, where the five county boroughs make up the city—we combine them into a single geographical unit. Although in some cases counties include areas outside of the city boundary, this approach also avoids the numerous boundary issues created by selecting census block groups based on census place boundaries. For example, Los Angeles and Detroit both include census places within their boundaries that despite being seamlessly integrated in the city infrastructure, would nevertheless be excluded.

Mobility Data

For our aggregate measures of mobility, we focus on the 25 largest cities in the United States. Mobility data comes from the SafeGraph Weekly Patterns data set. These data provide business locations with weekly information about the number, timing, and origin of visitors. As such, it includes daily counts of visits to over 5.6 million points of interest (POIs) and the home originating census block groups (SafeGraph 2022). SafeGraph derives this location data from a large sample of smartphone users.

As in any such sample, there are some concerns about how representative the sample is of the population at large (Coston et al. 2021; Li et al. 2023; Pew Research Center 2021). To mitigate potential demographic bias in the users captured by SafeGraph data, we aggregate data from the POI level and census block group level to the census tract level, where demographic bias is less of a concern (Li et al. 2023). In their multiscalar analysis of bias over time, Li and colleagues (2023) generally found that this aggregation process increases the correlation between population counts and the device population used to create the mobility networks. Our own replication of these analyses that arrive to similar conclusions are available in the Supplemental Information section “Bias in SafeGraph Data.” Thus, our final mobility data are a weekly network with census tracts as the nodes and edges weighted by the number or percentage of connecting trips as specified by the measures we introduce later. These data approximate and extend throughout 2021 and 2022 the widely used Social Distancing Metrics previously provided by SafeGraph to study mobility impacts of the COVID-19 pandemic (Chang et al. 2021; Gao et al. 2020; Glaeser et al. 2022; Levy et al. 2022; Marlow et al. 2021; Parolin and Lee 2021).

Demographic Composition and Other Place-Based Data

We supplement the mobility data with U.S. Census Bureau data. Population and demographic data are from the 2015 to 2019 American Community Survey (U.S. Census Bureau 2020). Given that our analysis is mostly focused on the connections between places, we also hone in on the differences in demographic composition between two places. We operationalize these differences in terms of their racial and ethnic composition and income composition. For income, we include 16 categories of yearly household income ranging from less than $10,000 to greater than $200,000. To measure race and ethnic composition, we include the percentage of the tract’s population that is Hispanic, non-Hispanic who identify as White, Black or African American, American Indian or Alaskan Native, Asian, Native Hawaiian and other Pacific Islander, some other race, and finally, individuals who identify as two or more races. Taking these values as the composition of the tract, we then use the Jensen-Shannon divergence (JSD) to measure the difference in the demographic distributions between the tracts of a dyad. We acknowledge that census data may not reflect the actual composition of tracts at the time of analysis, because these were collected prior to the study period, but data are unavailable for 2022, and the 2020 to 2021 data are incompatible with the older census geography SafeGraph uses.

Methods: Measuring Mobility Inequality

Our analysis centers on measuring aggregate changes in city mobility patterns. For this, we employ three previously developed indices of mobility network structure (Candipan et al. 2021; Phillips et al. 2021): the Concentrated Mobility Index (CMI), the Equitable Mobility Index (EMI), and the Segregated Mobility Index (SMI). In addition to these measures, we introduce a measure we call the Dissimilar Mobility Index (DMI) based on the sociodemographic dissimilarity between neighborhoods. Each of the four measures captures a distinct dimension of mobility network structure and thus should be thought of as complementary to one another.

The CMI measures the degree to which mobility in a city moves through a small number of “mobility hub” neighborhoods. The CMI is calculated as the Gini coefficient of the distribution of incoming trips at the neighborhood level. Neighborhoods with especially high numbers of visits skew the distribution and lead to larger values of CMI. Next, the EMI measures the distinct concept of “neighborhood isolation” by evaluating the degree to which weekly trips connect neighborhoods. In a perfectly even network (i.e., where no nodes are isolated), all neighborhoods are equally connected by trips. The EMI captures neighborhood isolation by measuring how much the observed mobility network deviates from an even one. It does this by first calculating the proportion of observed trips from a given census tract attributed to every other census tract. The resulting vector of values is then compared using the Hamming distance to an idealized vector of values where trips are evenly distributed. To scale the EMI between 0 and 1, the sum of Hamming distances for each neighborhood is divided by the theoretical maximum given the size of the network. Higher values of the EMI indicate a more even network, whereas values approaching zero indicate greater neighborhood isolation. Neither the EMI nor the CMI include any demographic information about neighborhoods and thus do not directly capture any information about the relationship between residential segregation and daily mobility. The final measure, the SMI, fills this gap by measuring the isolation between neighborhoods classified into their majority racial or ethnic group. To calculate the SMI, neighborhoods are categorized as majority Black, majority White, majority Hispanic, or as having no majority group. Mobility is then aggregated to capture the connections between these four groups. The final SMI value is calculated using the Hamming distance in the same way as the EMI. Unlike in Candipan et al. (2021), where the SMI was developed, we subtract the calculated value of the SMI from 1 so that like the EMI, smaller values indicate greater segregation and larger values indicate less segregation in the mobility network.

The SMI makes two simplifying assumptions that are important when interpreting its values. First, it categorizes the demographic composition of neighborhoods into belonging to one of only three demographic groups (White, Black, and Hispanic). Second, it reduces the mobility network to only connections among these and a fourth no-majority group. The concern is that both of these assumptions lose significant “lower level” information about demographic and mobility variation that may be consequential for interpreting the findings of the analysis of the SMI. For example, in many cities, mobility hubs that receive many daily visits are located in predominately affluent White central business districts. It is therefore possible that by reducing the classification of neighborhoods to four groups, the SMI will be sensitive to changes in the importance of mobility hubs.

The DMI

In an effort to maintain sensitivity to lower-level patterns in the data, we introduce a complementary measure, the DMI. Key to this index is our use of the Jensen-Shannon divergence (JSD) to calculate the social dissimilarity between neighborhoods. The JSD is one of several information-theory-based measures designed to quantify the difference between probability distributions (Cha 2007; Lin 1991). In addition to being widely used for machine learning applications, social scientists have used the JSD to measure phenomena such as the linguistic similarity of texts (Srivastava et al. 2018). For our purposes, the JSD has several appealing features. First, unlike a threshold classification based on the majority census group, the JSD can be calculated for probability distributions with an arbitrary number of categories. This allows us to incorporate a broad range of demographic characteristics into a single distance metric without making further simplifying assumptions about the distributions themselves. In our case, we take advantage of this flexibility by quantifying the difference between census tracts in terms of demographic composition by including a distribution of eight racial and ethnic groups or the full set of binned income frequencies provided by the U.S. census to create a race/ethnicity JSD and an income JSD, respectively.

To illustrate the usefulness of the JSD, consider the example ego network shown in Figure 1. Here, the origin neighborhood (O) is connected to four destination neighborhoods with varying distributions of four population groups (A, B, C, E). The first thing to note is that all units in this network have the same majority group (Group A). Thus, a representation of this network using a majority threshold would include four connections between identical neighborhood types. However, the distributions of the minority groups (Groups B, C, and E) vary across the other destinations. The JSD for each origin destination pair $(JSD(O,D_i))$ captures these differences continuously. For example, the origin and Destination 1 ( $D_1$ ) differ in which minority group is present. The resulting JSD is $0.4$ . Similarly, Destination 2 ( $D_2$ ) has two minority groups present and is therefore more similar to the origin, captured by a smaller JSD of $0.125$ . Furthermore, when the compositions of the tracts are identical, the JSD is equal to $0,$ as in the case of Destination 3 ( $D_3$ ). In the real world, these differences are likely meaningful. For example, a majority White tract with a large minority Hispanic population is probably meaningfully different from a majority White tract with a large minority Black population.

Figure 1.

Representation of a simple ego network for a neighborhood and the calculation of the JSDs. All nodes in this network have a majority group; however, their minority proportions differ, and thus, the JSDs vary across the edges. JSD = Jensen-Shannon divergence.

One caveat of the JSD is that it does not distinguish between the type of distributional differences across destinations. The JSD captures these differences only relative to the population of the home census tract. The relationship between the origin and Destination 4 ( $D_4$ ) illustrates this point. In $D_4,$ Group E is the second largest population group rather than Group C, as in $D_1$ . Despite this difference between the destinations (i.e., a JSD $=0.400)$ , their difference relative to the origin is identical, and so both connections have a JSD of $0.400$ . Although, if one were to compare the destinations to one another, their JSD would also be $0.400$ .

To calculate the JSD between a pair of tracts, we create a vector $O$ , whose elements define the probability distribution for the origin tract (i.e., the proportion of the eight groups in the tract), and a vector $D$ , whose elements define the probability distribution for the destination tract. We then calculate the JSD as:

$$JSD\left(O,D\right)=\frac{1}{2}KLD\left(O,M\right)+\frac{1}{2}KLD\left(D,M\right),$$

where $M=\frac{1}{2}\left(O+D\right)$ and $KLD$ is the Kullback-Leibler divergence, given as:

$$KLD\left(O,M\right)=\sum _c^C{O}_c{\log }_2\left(\frac{O_c}{M_c}\right),$$

where $c$ is an element of the vector $C$ , which defines the probability distribution of a tract over the demographic groups. In our case, we define $C$ in two ways. It either contains the distribution of incomes in a census tract or the proportion of a tract’s population that is Hispanic, non-Hispanic White, Black or African American, American Indian or Alaskan Native, Asian, Native Hawaiian and other Pacific Islander, some other race, and individuals who identify as two or more races. The base two logarithm produces a metric ranging from 0 to 1. Values closer to 1 indicate greater dissimilarity, and values of 0 indicate identical distributions.

Next, to measure dissimilarity within a city’s mobility network, we calculate the weighted average of the JSD. For each set of directed tract dyads, we weight the JSD by the proportion ( $p$ ) of all trips leaving the origin tract directed toward a destination tract. This approach assigns dyads that send and receive a larger proportion of trips’ larger weights. This way, although we do not introduce a threshold for two tracts to be “connected” as a function of the number of trips between them, those with few trips will contribute to this calculation negligibly. Networks, where mobility concentrates among tracts with similar compositions, will approach a value of 0, and networks with stronger interconnection between dissimilar places will have values approaching 1.

Finally, to increase the interpretability of the metric, we normalize using the unweighted average of all possible dyad JSDs. This unweighted average represents a mobility network where trips are evenly spread to all network nodes and is the total dissimilarity in the network. Thus, small values of the DMI indicate that mobility in the network is more concentrated between places with similar demographic compositions than an even network. Conversely, larger values of the DMI indicate that mobility is less concentrated between places with similar demographic compositions. Together, the equation for the DMI in a given city is simplified as:

$$DMI=\frac{1}{E}N{\sum }^{}JS{D}_{ij}*p_{ij},$$

where $N$ is the number of nodes, $JS{D}_{ij}$ is the JSD between census tract $i$ and $j$ , $p$ is the proportion of travel from tract $i$ directed toward tract $j$ , and the total dissimilarity in the network E, is given as:

$$E={\sum }^{}JS{D}_{ij}.$$

Functionally, the DMI is similar to information-based measures of residential segregation, such as the H-index (Reardon and Firebaugh 2002). Like the H-index, one appealing feature of this approach is that for each unique origin in the set of dyads, we can calculate the weighted average of the JSD to get a node-level measure of dissimilar mobility. This is especially useful because it enables multiscalar mapping and visualization of variation in the mobility networks across a city similar to other decomposable measures of segregation.

To illustrate the property of the DMI to capture node-level patterns, Figure 2 shows the mobility weighted average JSD of census tracts in Chicago, Houston, and San Francisco. These cities are geographically diverse and vary in their DMI. Compared to other cities, Chicago has one of the lowest observed DMI scores (0.64), indicating that tracts are more connected to more demographically similar places. Houston is close to the average DMI for places (0.75), and San Francisco is among the highest (0.85). Figure 2 shows that a clear spatial pattern emerges, with high average JSDs and low average JSDs tending to cluster together across urban space. Similar spatial patterning emerges in all three cities. Furthermore, the Moran’s I measure of spatial autocorrelation confirms that this visually evident spatial patterning is statistically significant (p < .001) in all cities (see Supplemental Information Figures S12–S14).

Figure 2.

Mobility weighted average JSD of census tracts in Chicago, Houston, and San Francisco. Mobility weights are based on travel during the week of April 15, 2019. Tracts with low estimated populations are not shown. JSD = Jensen-Shannon divergence.

One interesting thing to note about Figure 2 is that it is not straightforward to infer which cities will have a relatively low or high DMI. One might guess, for example, that because San Francisco appears to have relatively low average JSD scores at the tract level, it might also have a low DMI. Similarly, Chicago has the tracts with the highest mobility weighted average JSDs and thus might also be expected to have higher DMI. However, DMI is the ratio of observed mobility dissimilarity to the total dissimilarity among census tracts. Therefore, because San Francisco has a less diverse population, it also has a low level of overall dissimilarity. The DMI then indicates that a higher proportion of this low dissimilarity city is observed during daily mobility.

Table 1 presents the correlations among all four indices in the week of April 15, 2019. These correlations reveal interesting variation. First, we note that the SMI is not strongly correlated with either the DMI (r = .33, p $=0.11$ , df = 23) or the EMI (r = .17, p $=0.41$ , df = 23). This suggests that the SMI is capturing a distinct dimension of segregated mobility from the other measures. On the other hand, the EMI is strongly correlated with the DMI based on both race and ethnicity composition (r = .77, p < .001, df = 23), and income distributions (r = .62, p < .001, df = 23). These strong correlations suggest that less interconnected (i.e., more isolated) neighborhoods are also less demographically diverse. Finally, the CMI is not strongly correlated with any other measures. This makes sense because the CMI should capture a distinct dimension of a mobility network—the importance of mobility hubs in the overall network. Although further work is needed to understand the sources of differences among the measures, this is evidence for now that each captures separate dimensions of isolation and segregation in a mobility network.

Table 1.

Correlations between Mobility Indices for the Week of April 15, 2019.

	EMI	CMI	SMI	Race DMI	Income DMI
EMI	1.00	—	—	—	—
CMI	–0.10	1.00	—	—	—
SMI	0.17	–0.05	1.00	—	—
Race DMI	0.77***	0.13	0.33	1.00	—
Income DMI	0.62***	0.09	0.11	0.78***	1.00

Note: Other weeks in 2019 produce qualitatively similar results. EMI = Equitable Mobility Index; CMI = Concentrated Mobility Index; SMI = Segregated Mobility Index; DMI = Dissimilar Mobility Index.

***

p <.001.

Results

To start, we evaluate whether the previously documented changes to the structure of mobility networks that emerged in 2020 continued into 2021 and 2022. Figure 3 presents the average of a sample of 25 cities’ weekly measures of five indices characterizing distinct aspects of a weekly mobility network. ¹ For all measures, we calculate weekly values relative to their prepandemic 2019 values by subtracting the corresponding week’s 2019 value and dividing this difference by the 2019 value. Thus, all values shown in the descriptive figures represent the percentage change in the index value compared to 2019 levels. Values below 0 indicate a decline relative to the same week in 2019.

Figure 3.

Weekly average of five measures of mobility inequality 2020 to 2022 compared to their 2019 levels: the Concentrated Mobility Index (CMI), the Equitable Mobility Index (EMI), the Segregated Mobility Index (SMI), and the Dissimilar Mobility Index (DMI) in the 25 largest cities in the United States. The bands indicate 95 percent confidence intervals of the mean estimate of each index from the sample of cities.

Beginning in March 2020, all indices saw a significant decline relative to their 2019 baselines as mobility restrictions and concern over COVID-19 took hold nationwide (for weekly 2019 estimates, see Supplemental Information Figures S5 to S6). Here we show for the first time that the SMI follows a similar pattern: Mobility restrictions increased the racial and ethnic segregation in daily mobility networks by an average of 8 percent. This is a significant decline; compared to the same five-week period in March to April of 2019, the average weekly SMI estimates changed by 1 percent or less. By the end of 2020, the CMI and the SMI had returned to patterns resembling their pre-pandemic levels. Meanwhile, the EMI increased during the summer of 2020 but trended downward again as a new wave of infections spread across the country in late fall 2020.

The question we answer then, is whether neighborhood mobility patterns remained more isolated as indicated by the EMI or returned to pre-pandemic patterns bolstered by the decline in cases and the availability of vaccines. We find that although the EMI generally trended upward for all of 2021, it never fully returned to pre-pandemic levels. This means that even as pandemic restrictions were lifted and vaccines became widely available, neighborhoods remained relatively isolated from one another compared to 2019. Even more surprising is that the EMI remained lower than its 2019 baseline throughout 2022 as well, showing no sign of returning to previous patterns. The EMI was only briefly within 10 percent of its 2019 baseline in 2022, before declining and hovering around an average decline of 16 percent in the second half of the year. This finding is all the more surprising given the trajectory of the other mobility indices.

To better understand the sustained increase in neighborhood isolation starting in 2020, Figure 4 evaluates tract-level change between 2019 and 2022 in the mobility of three cities. We take advantage of the fact that the EMI is the scaled sum of tract-level Hamming distances (HDs) measuring the difference between the observed proportion of travel to each tract and a completely even mobility network. The first row of Figure 4 visualizes the difference in the tract-level HD values between April 2019 and April 2022 in Chicago, Houston, and San Francisco. Larger values of tract HD indicate greater differences from an even network and thus greater isolation. The second row of Figure 4 visualizes the significant values of the local indicator of spatial autocorrelation (LISA) based on a first-order queen contiguity spatial weights matrix (Anselin 1995). These maps help identify significant spatial patterning in values. Dark red indicates significant (p < .05) spatial clusters of relatively high HD values next to other relatively high values. Dark blue indicates clusters of low HD values next to other low values. The lighter colors, then, indicate places where relatively extreme values in either direction border areas with the opposite extreme (i.e., High-Low or Low-High). The sustained increase in neighborhood isolation through 2022 shown in Figure 3 is also evident in the first row of Figure 4. Tract-level HDs are slightly larger on average in 2022 than in 2019. This indicates that tract-level mobility is more different from an even mobility network in 2022 than in 2019. Furthermore, the lack of clear spatial patterning suggests that the changes in isolation are broadly experienced in each of the cities. Although this is true, the LISA maps in the second row of Figure 4 show how small variations do tend to cluster together into significant localized areas of increases (dark red) and decreases (dark blue) in mobility isolation relative to 2019 patterns. Further research is needed to understand the social and built-environment correlates of these spatial clusters. Our descriptive evaluation tentatively suggests that sustained increases in isolation were more common in dense urban areas of Chicago and Houston. Previous work suggests that this too could be linked to differences in car ownership and public transportation (Marlow et al. 2021).

Figure 4.

Change in tract-level Hamming distance values for three cities between April 2019 and April 2022. The first row shows the raw change in tract-level values. The second row visualizes clusters using a local Moran’s I (local indicator of spatial autocorrelation) statistic to identify statistically significant (p $<.05$ ) clusters of increases or decreases in values. Dark red indicates clusters of high values (High-High), dark blue indicates clusters of low values (Low-Low), and lighter colors indicate places where extreme values border the opposite extreme (High-Low and Low-High).

Other measures of mobility network structure differ from the EMI. Continuing their trends from the end of 2020, the CMI and the SMI values remain close to their 2019 baselines throughout 2021. The fact that the measure of segregation in the network (SMI) changes more similarly to the CMI than EMI is one surprising finding from this descriptive analysis. Given high residential segregation levels, we expected that changes in neighborhood isolation (EMI) would correlate with changes in the segregated mobility measure (SMI). However, this did not seem to be the case for most of 2021, and it was not until 2022 that changes in the CMI and the SMI significantly diverged for the first time. In 2022, the CMI declines and remains lowered so that in the second half of the year, the CMI is on average 3 percent lower than in 2019. This is a small but significant decrease compared to the 2019 baselines, which fluctuate by less than 2 percent over this time period (see Supplemental Figure S5). The SMI, on the other hand, increases slightly so that in the second half of the year, it is 2 percent higher than 2019. However, this small increase is not outside the estimated range of values over the same period in 2019.

The DMI complements the SMI by capturing the relative changes in a mobility network’s observed demographic dissimilarity (see Supplemental Information Figures S10 and S11 for the DMI alone). Overall, the income and race DMIs exhibit changes in 2020 and 2021 similar to the other indices shown in Figure 3. Starting in March 2020, the race and income DMIs decline rapidly during the initial phase of the pandemic, indicating that mobility became increasingly concentrated in neighborhoods more similar to home in terms of race, ethnicity, and income composition. By July 2020, both measures began to recover and continued to do so consistently throughout 2020 and 2021. Significantly, unlike the SMI and CMI measures, in 2020, the average of neither the income nor the race DMIs recover to 2019 values. It was not until the very end of 2021 that the estimated average DMI of cities reaches pre-pandemic levels. In this way, the DMI behaves more similarly to the EMI. However, in 2022, things change again. After an initial dip (seen also in the EMI), both measures of DMI increase throughout 2022. Similar to the SMI, the estimated mean of DMI in all 25 cities became positive in the second half of the year. However, there is a lot more variation in the DMI measures than in the SMI, and the confidence intervals rarely exclude 0. Therefore, although the upward trend may be suggestive, the DMI is still relatively close to the 2019 baseline.

Together, the EMI and the DMI indicate that structural changes initiated by lockdown procedures remained in 2021 and made neighborhoods more isolated from one another. These changes also show that neighborhoods had on average fewer connections to more demographically dissimilar neighborhoods. However, the DMI suggests this dissimilarity on average faded in urban mobility networks in 2022. The CMI and the SMI, on the other hand, suggest that with the recovery of central business districts and other mobility hubs, macro-level trends in segregated mobility recovered to their pre-pandemic patterns relatively quickly in 2020. In this way, our findings suggest that these indices work well together to understand multiple dimensions of mobility network structure.

These findings have some shortcomings that limit their generalizability. Although SafeGraph’s mobility data have been used frequently and found to be comparable to other mobility data, such as data provided by Google, Descartes, and the census’s own commuting tables (Chang et al. 2021; Kang et al. 2020), there are lingering concerns about the representativeness of the panel by demographic groups. On the one hand, SafeGraph reports a high correlation between their panel and census data, especially at higher levels of spatial aggregation (Li et al. 2023; Wang et al. 2021). Certainly, the pandemic created new temporary conditions that may be hard to detect. Specifically, for many Americans, the pandemic meant a loss of income and made paying for mobile phone subscriptions especially difficult for low-income individuals (Mcclain 2021). Although industry and survey research shows that mobile phone carriers reported a general increase in subscriptions over all quarters in 2020, with as many as 89 percent of urban residents owning a smartphone (Orf 2022; Pew Research Center 2021), we cannot rule out that short-term declines in certain groups might have meaningfully impacted the sample of users over particular time periods. We provide an evaluation of our sample in the Supplemental Information section “Bias in SafeGraph Data.” Like Li et al. (2023), we found that the sample of devices in the SafeGraph data was highly correlated within census block groups week to week (r $=.96$ ), there are no evident trends in the sample rate over the four years (see Supplemental Information Figure S15), and there are relatively low levels of demographic bias. In fact, African Americans are slightly overrepresented in the data, although some low-income groups and Hispanic people are slightly underrepresented. Looking at the correlations between tract populations reported in the American Community Survey (ACS) data and the device counts residing in a tract, however, we show that the correlation declines over time (see Supplemental Information Figure S17). This highlights another shortcoming of our overall approach, specifically, that relying on 2015 to 2019 ACS estimates for an analysis that spans 2019 to 2022 might introduce some bias over time. Unfortunately, this is a shortcoming we cannot overcome due in part to the lack of ACS data for all years but also because SafeGraph continues to issue data using 2010 census block boundaries, whereas the census has shifted to new geographic boundaries without a way to completely accurately translate between the two.

Conclusion

Our analysis of weekly mobility patterns indicates that some of the structural changes to neighborhood mobility networks that emerged in 2020 in response to the COVID-19 pandemic remained until the end of 2022. In particular, we show that neighborhoods in the 25 largest U.S. cities remained more isolated from one another throughout 2020 to 2022 compared to 2019. Similarly, we found evidence that weekly travel in 2020 and 2021 was, on average, concentrated in places more similar in terms of their racial, ethnic, and income composition. These findings are significant because they suggest that the changes to mobility initiated by the response to the COVID-19 pandemic in March 2020 had durable impacts on dimensions of mobility, which connect to other types of social inequality (Cagney et al. 2020; Levy et al. 2020).

Furthermore, we contribute methodologically to the growing literature on urban mobility (Wang, Zhang, and Li 2022) by introducing a new measure of segregated mobility that we call the Dissimilar Mobility Index (DMI). Due to its use of the Jensen-Shannon divergence as a continuous measure of the difference between any two probability distributions, the DMI is sensitive and flexible in the types of mobility segregation it can measure. The DMI is also decomposable into node-level measures, allowing for multiscalar analysis of segregated mobility patterns. We demonstrate this flexibility by evaluating income and racially segregated mobility over 2019 to 2022. We show that despite a different approach to measurement, the DMI is related to the isolated mobility measure, the EMI, and displays characteristics of a longer-lasting alteration after March 2020. Specifically, city mobility networks, on average, became more racially and income-segregated throughout 2020 and 2021. However, unlike the EMI, the DMI did return to 2019 levels in 2022, even as isolation remained higher than in 2019.

The present study does not investigate the correlates of these changes over time; therefore, we highlight avenues for future research. In particular, the changes observed in 2022 especially call out for additional analysis. What drove the decline in the importance of mobility hubs in 2022? And why did the measures of segregated and dissimilar mobility increase? Although COVID-19 remained ever present during this time period, economic issues, such as inflation, emerged in 2022 that may be important for understanding observed changes in mobility. Future work should engage with these important questions, looking to study alternative sociodemographic distributions within census tracts or apply community detection and regionalization methods to detect alternative spatial scales of mobility dissimilarity (Chodrow 2017; Huang et al. 2022; Vachuska 2023).

Taken together, our findings contribute to the growing sociological literature on the nature of urban mobility and its relationship to residential segregation. Furthermore, our results help us better understand how these relationships evolved during the COVID-19 pandemic in ways that may have long-lasting consequences.

Supplemental Material

sj-docx-1-srd-10.1177_23780231231198857 – Supplemental material for Durable Change in U.S. Urban Mobility Networks, 2019–2022

Supplemental material, sj-docx-1-srd-10.1177_23780231231198857 for Durable Change in U.S. Urban Mobility Networks, 2019–2022 by Thomas Marlow, Kinga Makovi and Bruno Abrahao in Socius