The contingency of neighbourhood diversity: Variation of social context using mobile phone application data

Abstract

Chinese

This research uses high-density anonymised mobile phone application (MPA) global-positioning system (GPS) data to describe exposure to racial diversity in different social contexts with an aim to clarify the mechanism linking residential diversity to opportunities for diverse social interactions. In particular, it explores the hypothesis that a diverse residential context does not lead to diverse social contact by comparing three exposure measures – residential, observed and interaction – on the census block group level in Chicago. In doing so, it also explores the contribution of activity spaces to opportunities for diverse social contact. The findings show that the exposure to opportunities for diverse social contact measured by MPA data is generally higher than what is implied by residential census data, especially in areas of high residential segregation in the city. Further, measures using MPA data reveal more spatiotemporal heterogeneity of exposure than that implied by the residential context.

Keywords

big data GPS data neighbourhood diversity racial segregation social interaction

Introduction

The proposal that exposure to neighbourhood diversity improves socioeconomic outcomes through social and physical integration is central to current theory and practice in urban development and housing policy in North America. It is seen as a response to the deleterious consequences of socioeconomic and racial segregation at the neighbourhood level, especially in the context of residential segregation (Freeman, 2009; Yinger, 1986), which can mediate social processes and amplify the concentration of poverty (Crowder et al., 2012; Wilson, 1987). Neighbourhood diversity is also the central aim of many affordable housing measures (Goetz, 2003; Schwartz, 2014). Its efficacy, however, is often challenged on the insufficient empirical grounding to explain the mechanism that links diversity of context to actual diverse social exposure and contact (Galster, 2012; Galster and Friedrichs, 2015).

Previous research on housing mix suggests diverse social contact is less than expected given neighbourhood demographics (Blokland and Van Eijk, 2010; Butler, 2003; Robson and Butler, 2001). While these studies provide meaningful insight into the social processes and networks that compose neighbourhoods, they are often case studies or ethnographic studies that may be contingent on particular social, political and institutional contexts. Quantitative methods used to describe neighbourhood diversity generally rely on census demographics, presenting an epistemological challenge as census data typically only reflect the residential context despite the diversity of activity spaces – the physical locations individuals have direct contact with as a result of day-to-day activities (Golledge and Stimson, 1997; Horton and Reynolds, 1971; Lynch, 1960). This methodological divide potentially abets contradictory conclusions and theoretical divisions.

This analysis uses high-density anonymised MPA GPS (simplified to ‘MPA’ from here) data to detect activity spaces and test whether exposure to racial diversity in a neighbourhood differs across three types of exposure measures, which I consider to delineate three increasingly representative opportunities of social interaction: residential exposure as measured by census data, observed exposure as measured by all observed activity in a neighbourhood and interaction exposure, which measures the diversity of opportunities for interactions as defined by a space-time activity space overlap of two or more people in the MPA data set. Tracing the links and differences between these three measures can clarify the mechanism by which residential and observed social context translate into opportunities for social interaction.

This study seeks to investigate the following research questions: Can MPA data be representative of the urban population? How do the three exposure measures differ in their estimated exposure to racial diversity? If actual diverse social contact occurs less often than presumed in a diverse neighbourhood, as suggested by previous literature, is this evident in how interaction exposure behaves relative to residential and observed exposure? I hypothesise that:

The MPA data can be used to sufficiently replicate Chicago’s residential population.

Observed exposure to racial diversity will be higher than residential exposure as it represents a wider range of activity spaces within the neighbourhood.

Interaction exposure will be lower than residential exposure, as previous literature suggests that actual social interaction is less diverse than residential context might indicate.

If hypothesis (1) can be validated, it can enable us to test how these three different types of social contexts relate to one another. A wider range of activities and opportunities of encounter should be considered for a more accurate characterisation of segregation dynamics beyond the residential context and MPA data may be sufficiently robust in characterising these contexts. Though the analysis remains at the level of measuring opportunities for social interaction rather than actual contact, I suggest that the likelihood of contact is a function of physical proximity. Developing a better understanding of (2) and (3) will shed light on the gap between our theoretical and empirical knowledge of social interaction mechanisms and their relation to neighbourhood diversity.

Background

Exposure to neighbourhood diversity

The neighbourhood context has material and social influences on residents in terms of access to education, health, public and private services, crime and employment opportunities. Previous research has shown that residential location encodes spatially manifested inequities such as exposure to particular environmental conditions (Bowen et al., 1995), differential access to services (Altschuler et al., 2004), amenities (Lee and Lin, 2017), as well as opportunities (Galster and Killen, 1995). In the USA context in particular, the legacy of concentrations of poverty in racially segregated, lower-income neighbourhoods (Cutler et al., 2002; Wilson, 1987) has motivated measures to deconcentrate poverty and promote social integration, typically enacted through housing (Briggs, 2005; Galster, 2002). Studies suggest that improved economic opportunities (Cutler et al., 2008), intergenerational prospects (Chetty et al., 2016), social trust (Abascal and Baldassarri, 2015; Putnam, 2007) and opportunities for professional mobility (Galster et al., 2008) are all associated with neighbourhood diversity.

Understanding the social contact mechanism

The primary mechanism through which neighbourhood diversity is theorised to improve social outcomes is a link between increased exposure to diversity and increased diversity of social interactions (Galster and Friedrichs, 2015). The effects of this link have not always been clear or consistent, or they suggest only small benefits (Manley et al., 2011; Ostendorf et al., 2001). For example, Crowder et al. (2012) find that, despite the growing prevalence of multiethnic ‘global neighbourhoods’ (Logan and Zhang, 2010), social integration and mobility remain stratified along racial lines.

Galster and Friedrichs (2015) propose that the diverse social contact implied by neighbourhood diversity is rarely present. Previous literature suggests diverse neighbourhoods may not be conducive to actual exposure to diversity. True social mix is, at best, mediated by race, class and ethnicity divisions (Blokland and Van Eijk, 2010; Galster et al., 2010; Robson and Butler, 2001), social network distance (Arthurson, 2012), or different time schedules (Fraser et al., 2013). Robson and Butler (2001) describe the ‘tectonic’ nature of the parallel and non-integrated social relations within gentrifying neighbourhoods that are purportedly in favour of diverse racial and ethnicity diversity. The incongruity between the theory and evidence begs clarification, with Galster (2002) suggesting the necessary and sufficient conditions for interaction may be more stringent than expected. Though confirmed interactions are not investigated here, this analysis isolates spatiotemporal overlaps of activity spaces as greater opportunities for interaction.

Measuring exposure to neighbourhood diversity

Beyond the residential context

Residential context as typically defined by census data and administrative boundaries has a sociological meaning; however, Kwan (2013) notes that the focus on residential census data is static and limits our scope in describing daily activities across a range of activity spaces. Residents often do not even perceive or consider administrative neighbourhood boundaries in activity decisions (Coulton et al., 2001, 2004). Revisions of traditional notions of residential diversity have developed a more ego-centric and temporally dynamic characterisation of the activity space (Boterman and Musterd, 2016; Browning and Soller, 2014; Candipan et al., 2021; Kwan, 2009, 2013; Matthews, 2011; Vallée et al., 2010; Wong and Shaw, 2011). In particular, Browning and Soller (2014) study routine social exposure to activity spaces and conclude that shared local exposure in activity spaces is conducive to higher levels of social capital. Similar to this paper’s questions, Jones and Pebley (2014) use granular survey data to ask whether residential neighbourhoods are representative social contexts. They find the activity space landscape more heterogeneous than residential contexts.

Using mobile phone GPS data to understand neighbourhood diversity

Anonymised MPA data represent actual activities across time and space, including but not limited to the residential one. Large geographic coverage of mobile phone data and the possibility of scaling mobility counts can facilitate comparison across different regions. Previous studies using call detailed record (CDR) data have shown the accuracy of population estimates against actual population counts (Deville et al., 2014; Jiang et al., 2017; Liu et al., 2018; Yao et al., 2017), which may be used to improve understanding of dynamic urban flows (Reades et al., 2007; Yuan and Raubal, 2012) and the feasibility of using these data to represent individual activity spaces (Calabrese et al., 2013; Iovan et al., 2013; Jiang et al., 2017). Whereas CDR data have a coarse accuracy of around 200–300 m (Jiang et al., 2013), MPA data have greater accuracy, with a median accuracy of 12 m in this paper.

Other recent studies using GPS data have demonstrated racialised or class-based social-spatial isolation (Palmer et al., 2013; Phillips et al., 2019; Wang et al., 2018), the importance of non-residential contexts in understanding exposure (Jones and Pebley, 2014; York Cornwell and Cagney, 2017), and the spatiotemporal heterogeneity of exposure even for residents of the same neighbourhood (Browning et al., 2017a).

Data and methods

This study analyses three types of exposure to racial diversity on the census block group level: residential exposure as implied by the 2014–2018 American Community Survey (ACS) demographic data, observed exposure representing the empirically measured stays in the neighbourhood using MPA data, and interaction exposure, also using MPA data, where individuals share a space-time overlap in activity spaces, creating opportunities for social interaction.

In this data set, there are approximately 40 million data points daily, with the analysis period between 1 July and 31 October 2019. The final data set includes home locations for 87,631 residents in the city with a total of 1,644,955 stays and 29,888 unique interaction clusters. While this ultimately represents 3.4% of Chicago’s 2016 population (as the representative population for the 2014–2018 ACS data) within the 2103 residential census blocks studied, with some potential selection bias, an analysis of data representativeness against the ACS population data (detailed in the following section) suggests an unbiased estimation of population can be created.

Data pre-processing

The use of anonymised MPA data is provided by Cuebiq, a location intelligence company that collects de-identified data from opted-in users,¹ takes advantage of both high resolution and temporal variation in the data in describing individual activity patterns. The data are derived from applications such as transportation, social media, exercise and weather categories, amongst others, within Chicago city boundaries. The characteristics in the data set are: latitude and longitude, location accuracy (in metres), timestamp and an anonymised and encrypted user ID.

A summary of the data cleaning, filtering and aggregation process is provided in Table 1, along with summary statistics. The initial stage of data pre-processing concerns removing data with poor accuracy and data on transit, because of the difficulty of distinguishing vehicular modalities – using this data. driving alone in a car, which is not conducive to exposure (Wong and Shaw, 2011), appears effectively the same as multiple people on a bus. All data that have higher than a 200 m accuracy – roughly the radius of a circle inscribed in an average block group area of 165 km² in the city – are removed. All points within a 18.29 m (or 60 ft) diameter buffer around the arterial road centrelines are also removed.² This diameter is determined by the typical residential street width in Chicago of 20.12 m (or 66 ft) diameter.³ All points with velocities greater than 24 km per hour or 14.91 miles per hour – faster than a standard walking pace but slower than the average vehicular speeds of 24 miles per hour in the city (Smith, 2018) – are also removed. This corrects for some of the selection bias as more than half the data appear to be transit data by these metrics.

Table 1.

Summary of data preprocessing, home and stay locations and interaction cluster creation.

Description	Resulting	Measure	Percentiles^a
Original data: Approximately 40 million points daily			25%	50%	75%
Pre-processing
1.Remove data with greater than 200 m accuracy	97% of original	Accuracy (m)	5	12	65
2.Remove data with greater than 24 km/hr velocity	43% of original	Velocity (km/h)	0.3	2.2	10.8
3.Remove all data with a 18.29 m buffer around major streets centrelines	38% of original	Velocity (km/h)	0.3	2.3	10.8
Home locations
4.Find potential home locations per user: for data across entire period, cluster user locations for a minimum of 3 points per cluster within a 150 m radius between 01:00 and 05:00 h. Return the centroid of each cluster	743,968 centroids	Count per user	1	2	5
5.Remove all home locations not in residential, downtown, or planned development districts	736,968 centroids	Count per user	1	2	5
6.Find the ‘best guess’ final home location: for users with more than one home location, all possible locations within a 150 m radius are clustered and the centroid of the highest count cluster is considered home location	129,997 people	Count per user	1	1	1
7.Remove all block groups with high MPA home density to actual population density ratio, based on resulting correlation of approximately 0.8	87,631 people	Count per user	1	1	1
Stay locations
8.For data across entire period, cluster user locations within a 10 m radius and 5-minute window, where each cluster has at least 3 points, to find ‘stay’ locations. Return the centroid of each cluster	2,456,050 centroids	Count per user per day	3	3	5
9.Remove all data with no home location for the user	1,644,955 centroids	Count per user per day	3	3	5
Interaction clusters
10.Cluster stay locations within a 10 m radius and 5-minute window. Return clusters with two or more people in each cluster	29,888 clusters	Count of users per cluster	2	3	4

Notes: ^aFor pre-processing, percentiles for raw data based on a representative day. For home and stay location, percentiles for location centroids per user across the entire period. For interaction clusters, percentiles for each interaction cluster.

Stays and demographic profiles

Two fundamental assumptions underpin the exposure measures using the MPA data: the first is the assumption of a probabilistic demographic profile for each individual in the data based on their estimated home locations (Wong and Shaw, 2011), which are then joined to census data. The second is the assumption that an individual’s daily activity space can be extrapolated from the data.

The initial step for the study is an estimation of home locations for each individual. This is estimated using the density-based spatial clustering of applications with noise (DBSCAN) algorithm (Ester et al., 1996). DBSCAN allows the discovery of activity spaces based on a density of raw points for a given radius. The process of estimating home locations is summarised in Table 1. An individual’s home location is defined as the centroid of their most frequented clusters, using the DBSCAN algorithm with a maximum 150 m threshold⁴ from the hours of 01:00 CT to 05:00 CT, when individuals are generally likely to be at home. The 150 m threshold is created to be inscribed within an average Chicago block group. These are then filtered by Chicago zoning districts to keep only those locations within feasible areas of residence according to zoning regulations (residential, downtown and planned development districts). Block groups that show a high ratio of home location densities versus the census population density – primarily in areas of high activity across all hours of the day such as downtown and both O’Hare and Midway airports – are also removed from this analysis, because of the higher probability these are not home activities. The details of this process are described in Table 1.

This analysis employs the notion of ‘stay’ locations (Jiang et al., 2013, 2017) as the centroid of a cluster of at least three different raw points that are within 10 m and 5 minutes of one another.⁵ Stay centroids are kept only for individuals who have home locations in Chicago city boundaries as a portion of the individuals (around 67%) are visitors to the city. These stay centroids (simplified as ‘stays’ from here on) operationalise the concept of activity spaces using MPA data. The higher precision of the MPA data allows for a smaller clustering radius that could realistically represent close physical proximity. From stays, it is then possible to distinguish spatial-temporal overlaps of activity spaces for multiple people – an opportunity for interaction. Like stays themselves, interaction clusters are determined by stays that are within 10 m and 5 minutes of one another for at least two different residents. Figure 1 illustrates an example of this.

Figure 1.

Example of how overlapping stays (shaded circles) create an interaction cluster (outlined in grey). Individuals A and B have stays that are within 10 m range from one another and also overlap in time.

Residential location and demographics

The resulting data represent a sample of 87,631 anonymous residents. Though the number of residents in this analysis is smaller than the estimated 300,000 total number of users in the original data set due to the necessity of filtering the data for residents, the resulting Pearson correlation of the MPA-derived log population density to the ACS log population density on the Census block group is 0.83 and is within the same range of other studies using similar scales (Bachir et al., 2017; Deville et al., 2014).

Derived home locations and other census characteristics are used to create an estimated population based on the 2014–2018 American Community Census 5-Year estimates at the block group level, along with probabilistic demographic profiles for each resident in the final data set. The estimated population, using the ordinary least squares (OLS) model below, is employed across all exposure calculations. For block group $i$ , $\log ρ_{i}$ is the log census population density, $\log σ_{i}$ is the log mobile phone population density, $γ_{i}$ is the mobile phone population, $x_{1 i}$ the percentage Black population and $x_{2 i}$ the percentage higher education are used as controls in the model to account for variations in coverage that may arise from socio-economic conditions.⁶ $α$ is a constant and $β_{1} \dots β_{4}$ are coefficients, all to be estimated.

\log ρ_{i} = α + β_{1} \log σ_{i} + β_{2} γ_{i} + β_{3} x_{1 i} + β_{4} x_{2 i}

(1)

This model produces an unbiased prediction of population density with an adjusted R-squared of 0.793,⁷ confirming Hypothesis 1. This model is then used to create a predicted population ${\hat{y}}_{i} = θ_{i} {\hat{ρ}}_{i}$ where $θ_{i}$ is the area of the block group. The predicted population is used to create weights $w_{i} = \frac{y_{i}}{{\hat{y}}_{i}}$ to correct for over- or under-representation for a particular block group’s demographics. For example: if the weight for a particular block group is 1.1, meaning that block group is under-represented in the predicted population, then any resident in the estimated population from this block group will represent 1.1 persons. The average value of these weights is 1.01, 0.82 at the 25th percentile and 1.25 at the 75th percentile. Figure 2 illustrates the estimated and actual demographic distributions for population, population density, percentage Hispanic, percentage non-Hispanic Black, percentage non-Hispanic White, percentage non-Hispanic Asian, percentage with a bachelor’s degree or higher and median income. Though there are slight discrepancies, the racial and ethnicity distributions used in this analysis almost entirely overlap.

Figure 2.

Distribution of predicted versus actual census demographics.

To mitigate ecological fallacies, each individual is assigned a probability of membership in a particular group based on their home census block demographics in the 2014–2018 ACS 5-Year estimates. For race and ethnicity, I use the percentage categories for non-Hispanic White, non-Hispanic Black, non-Hispanic Asian and Hispanic. For instance, if an individual’s home location is in a block group with 70% non-Hispanic Black, 20% non-Hispanic White, 5% non-Hispanic Asian and 5% Hispanic, their probability of membership in each racial category will be the same.

Exposure measure

Standard measures of exposure to diverse social contact are based on Lieberson’s (1981) formulation. This analysis uses a measure of local interaction potential based on the Simpson index – similar in nature to the standard Lieberson measure – which reflects the probability that two randomly selected people will belong to members of different groups and is agnostic to boundaries of area units and the number of groups in the comparison. The three exposure measures are calculations of the Simpson index in three different contexts. Here the baseline residential exposure for each block group $i$ is:

S_{i}^{A} = 1 - \sum_{m} {(\frac{{\hat{y}}_{mi}}{{\hat{y}}_{i}})}^{2}

(2)

where ${\hat{y}}_{mi}$ is the estimated population for each racial or ethnicity category $m \in M$ in block group $i$ and ${\hat{y}}_{i}$ is the block group’s total estimated population of all four race and ethnicity categories included in the analysis.⁸ All exposure measures have a range of $[0, \frac{M - 1}{M}]$ , since the second term of the measure is a subtraction of the self-exposure, or isolation, probability. Thus, any exposure measure can be a maximum of 0.75 in this analysis. Though the estimated population is used to calculate residential exposure, the within unit proportion of each racial group for $S_{i}^{A}$ remains the same as the original census data and no weighting is necessary. For observed exposure, because exposures are now calculated based on all stays within a block group, including those by residents not in the block group, the populations are the weighted average proportions for group racial category as each resident is now weighted:

\begin{matrix} S_{i}^{B} = 1 - \sum_{m} {(\frac{\sum_{j} w_{j} {\hat{y}}_{mj} / \sum_{j} w_{j}}{\sum_{j} w_{j} {\hat{y}}_{j} / \sum_{j} w_{j}})}^{2} \\ = 1 - \sum_{m} {(\frac{{\bar{y}}_{mi}}{{\bar{y}}_{i}})}^{2} \end{matrix}

(3)

where $j$ indexes on all home block groups of people that appeared in $i$ and $j$ can be equal to $i$ . $w_{j}$ is the population weight for each block group. Lastly, interaction exposure looks at the exposure in shared activity spaces, as operationalised by interaction clusters. For each shared activity space $l$ :

S_{l} = 1 - \sum_{m} {(\frac{\sum_{k} w_{k} {\hat{y}}_{mk} / \sum_{k} w_{k}}{\sum_{k} w_{k} {\hat{y}}_{k} / \sum_{k} w_{k}})}^{2}

(4)

where k indexes on clusters within i. All shared activity spaces within a block group $i$ are aggregated to a population-weighted average within that block group in order to facilitate comparison:

S_{i}^{C} = \frac{\sum_{l} w_{l} S_{l}}{\sum_{l} w_{l}}

(5)

Results

Table 2 summarises the results of comparisons between three exposure measures, with three different radii for interaction clusters to determine sensitivity of the interaction exposure to distance. The median for the residential exposure is 0.33, indicating that there is a 33% probability that any two random residents in the block group are of different races. The observed exposure probability is generally higher than the residential exposure, with a median of 0.49, which confirms Hypothesis 2. This result may be intuitive as the observed exposure represents a wider range of activities spaces, including the residential space implied by the residential exposure measure. Residential exposure is less than 0.13 for 25% of block groups – unsurprising given the city’s residential segregation along racial lines (Massey, 2016; Massey and Tannen, 2015; Sampson, 2011) – while the other measures are closer to 30% at the 25th percentile. The interaction exposure at the 10 m radius is 0.44 and only slightly less than observed exposure. The statistics for larger radii indicate that the exposure levels do not change much with a looser social interaction clustering radius. There is an 81% and 63% correlation between the residential exposure and observed and interaction exposures, respectively, implying that areas of high residential exposure are also areas with high exposure for other kinds of activity spaces. For the majority of block groups, the observed exposure is significantly higher than the residential exposure, as shown in Figure 3. In terms of the distribution of interaction exposures these are all as expectedly lower than the observed exposures, though surprisingly, most interaction exposure levels are higher than the residential exposure. In other words, when people have an opportunity to interact, the degree of exposure to racial diversity generally tends to be higher than what is implied by the residential exposure measures. Thus, Hypothesis 3 cannot be firmed.

Table 2.

Exposure score summary.

	Residential	Observed	Interaction
			10 m	20 m	30 m
Mean	0.32	0.45	0.41	0.40	0.40
Std dev.	0.21	0.17	0.18	0.18	0.17
Min.	0.00	0.01	0.00	0.00	0.00
25th percentile	0.13	0.33	0.28	0.29	0.28
Median	0.33	0.49	0.44	0.44	0.43
75th percentile	0.51	0.59	0.54	0.54	0.54
Max.	0.75	0.75	0.73	0.73	0.73
Pearson correlation to residential		0.81	0.63	0.69	0.72
Census block groups in sample	2103	2103	960	1429	1693

Figure 3.

Residential exposure (left), difference between observed and residential exposure (middle) difference between interaction and residential exposure at 10 m radius (right).

Figure 3 shows the geographic variation of residential exposure and how the observed and interaction exposure measures deviate from residential. All three exposure measures have generally similar regions of high and low exposure levels (in the middle and right maps in Figure 3 these are grey and light-coloured areas near 0). Regions that are most exposed to diversity are along the city’s major roadways where the city has seen growth and gentrification in the last few decades and where there are significant Hispanic populations. The middle map in Figure 3 shows that observed exposure levels are higher than residential exposure, especially in the city’s South and West sides, which suggests an underestimation of exposure when relying on census data. The rightmost map showing the difference between interaction and residential exposure loosely corroborates this proposition but also highlights the spatial heterogeneity of the interaction measure. Of note is the total number of census blocks covered by the interaction exposure measure, which is a smaller subset of the residential and observed measures. This is due to the fact that not all blocks contained interaction clusters; interaction exposure is conditional on individuals having the opportunity to come into close proximity.

Figure 4 shows the density of social interaction clusters and their average interaction exposure measure by the hour across the days of the week. The density of clusters are typically higher during the day and begin to decline around 16:00 h and also generally declining starting on Thursdays (with August being an exception, where density is higher starting Thursdays). The mean interaction exposures tend to be highest between the hours of 04:00 h and 12:00 on weekdays and gradually decline over the course the afternoon and evening, with exposures generally lower on the weekends.⁹

Figure 4.

Density distribution of social interaction clusters across the week by hour (top). Average interaction exposure throughout the week by hour, with shading representing the 95% confidence interval (bottom).

Discussion and conclusion

This analysis demonstrates the possibility of creating an unbiased and representative predicted population and also describes possible links between neighbourhood diversity and opportunities for diverse social interaction through comparing residential exposure to actual and more specific contexts of interaction opportunities. The results confirm Hypothesis 2: observed exposure to racial diversity is higher than residential exposure, with West side neighbourhoods such as North Lawndale and Garfield Park and South side neighbourhoods such Washington Park and Englewood showing much higher exposures than the residential data in these ‘hyper-segregated’ (Massey and Denton, 1989) neighbourhoods would indicate. Nevertheless, observed exposure is still fairly correlated with residential measures. Although here I consider the full scope of an individual’s daily activity space, this finding aligns with previous literature comparing workspaces and transit activities exposures to residential exposures.

The findings regarding Hypothesis 3 are more surprising. The interaction exposure generally shows higher values than residential exposure, conditional on opportunities for interactions occurring in space and time. This is likely also due to the wider range of activity spaces accounted for when using MPA data. The findings from interaction exposures also reveal a higher degree of spatial heterogeneity relative to residential exposure, even on racially segregated West and South sides of the city, which confirms previous research (Jones and Pebley, 2014). Of note are the pockets of high interaction exposure in these parts of the city next to ones of low exposure. Ultimately, these results do not necessarily refute previous research. The interaction exposure describes a set of opportunities for interaction and not actual interactions; further, in this paper, these opportunities are defined by space-time proximity. This may engender one type of opportunity for interaction but exclude other types of proximity that are also conducive to interaction, such as social network proximity. The weekly temporal variation of the degree of interaction suggests a possible source of exposure: exposure is generally higher during regular weekday activities, perhaps mostly consistently articulated in the morning, which are likely workplace or school activities. The monthly variation of exposure in Figure 4, especially the reduced interaction clusters in October, also suggests seasonality in these patterns and sensitivity of exposure due to the weather. Lastly, the degree of overlap across all three exposure measures suggests that residential context remains an important factor in determining exposures across a broader range of activities and contexts. This may result from residents extending their activity spaces most frequently to places in and near their home locations.

There are several limitations and future research questions that this analysis introduces: first, there are questions of representativeness of the home and stay locations. The estimation of the home locations assumes the MPA users represent a random sample of residents in the block group, which may not be the case if there is selection bias for MPA users. Furthermore, while the residential population can be estimated using the MPA data, the lack of reliable temporally dynamic population counts means temporal calibration of MPA stay data is difficult to validate. This presents a potential area of future research. Second, as previously mentioned, opportunities for interaction do not necessarily result in actual interaction. The limit in precision in data – 12 m at the median – also constrains more detailed analysis; thus, future research may explore mixed methods approaches to validating actual interactions. Third, future research could study whether exposure differs for different racial and ethnic groups while also considering its intersection with socioeconomic status (Browning et al., 2017b). Lastly, though use of these data for research purposes required compliance with privacy restrictions set by the provider, the increasing availability of MPA data raises broader questions of privacy protection. While the USA lacks legislation similar to the European Union’s General Data Protection Regulation, research on geographic differential privacy to obscure traceability is increasingly prevalent (Andrés et al., 2013; Xiao and Xiong, 2015) and should be considered in future analysis.

From a policy perspective, most affordable housing policies in the USA favour mixed-income neighbourhood diversity and ‘policy-driven gentrification’ approaches, despite the unresolved link between neighbourhood diversity and opportunity through, amongst other influences, the potential for social contact. This article suggests there is more potential exposure to diverse social contact than previous literature finds and opens up new methods to explore the social contact mechanism, which also demonstrates an empirical and scalable approach to elucidating these policy questions. Future investigations that incorporate specific policies, particular groups of interest and an intersectional approach to diversity, such as between race and class, may improve our understanding of the overall effects of social mixing-driven housing and urban development policy.

Footnotes

Acknowledgements

This article was originally presented at the ‘Predicting neighborhood change using big data and machine learning: Implications for theory, methods, and practice’ symposium on January 9, 2020 at the University of California Berkeley. Thank you to the organizers and attendees for their feedback on an earlier draft. I would also like to thank Lance Freeman, Jamie Saxon, Shin Bin Tan, the presenters and attendees of the Association of Collegiate Schools of Planning Annual Conference on November 5, 2020, David Weisberger, the anonymous reviewers, and the editor for their feedback.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Wenfei Xu

Notes

References

Abascal

Baldassarri

(2015) Love thy neighbor? Ethnoracial diversity and trust reexamined. American Journal of Sociology 121(3): 722–782.

Altschuler

Somkin

Adler

(2004) Local services and amenities, neighborhood social capital, and health. Social Science & Medicine 59(6): 1219–1229.

Andrés

Bordenabe

Chatzikokolakis

, et al. (2013) Geo-indistinguishability: Differential privacy for location-based systems. In: Proceedings of the 2013 ACM SIGSAC Conference on Computer & Communications Security. Berlin, Germany, 4 November. New York, NY: Association for Computing Machinery, pp. 901–914.

Arthurson

(2012) Social Mix and the City: Challenging the Mixed Communities Consensus in Housing and Urban Planning Policies. Collingwood: CSIRO Publishing.

Bachir

Gauthier

El Yacoubi

, et al. (2017) Using mobile phone data analysis for the estimation of daily urban dynamics. In: 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC). Yokohama, Japan, 16–19 October. IEEE, pp. 626–632.

Blokland

Van Eijk

(2010) Do people who like diversity practice diversity in neighbourhood life? Neighbourhood use and the social networks of ‘diversity-seekers’ in a mixed neighbourhood in the Netherlands. Journal of Ethnic and Migration Studies 36(2): 313–332.

Boterman

Musterd

(2016) Cocooning urban life: Exposure to diversity in neighbourhoods, workplaces and transport. Cities 59: 139–147.

Bowen

Salling

Haynes

, et al. (1995) Toward environmental justice: Spatial equity in Ohio and Cleveland. Annals of the Association of American Geographers 85(4): 641–663.

Briggs X de

(ed.) (2005) The Geography of Opportunity: Race and Housing Choice in Metropolitan America. Washington, DC: Brookings Institution Press.

10.

Browning

Soller

(2014) Moving beyond neighborhood: Activity spaces and ecological networks as contexts for youth development. Cityscape 16(1): 165.

11.

Browning

Calder

Ford

, et al. (2017a) Understanding racial differences in exposure to violent areas: Integrating survey, smartphone, and administrative data resources. The Annals of the American Academy of Political and Social Science 669(1): 41–62.

12.

Browning

Calder

Krivo

, et al. (2017b) Socioeconomic segregation of activity spaces in urban neighborhoods: Does shared residence mean shared routines? RSF: The Russell Sage Foundation Journal of the Social Sciences 3(2): 210–231.

13.

Butler

(2003) Living in the bubble: Gentrification and its ‘others’ in North London. Urban Studies 40(12): 2469–2486.

14.

Calabrese

Diao

Di Lorenzo

, et al. (2013) Understanding individual mobility patterns from urban sensing data: A mobile phone trace example. Transportation Research Part C: Emerging Technologies 26: 301–313.

15.

Candipan

Phillips

Sampson

Small

(2021) From residence to movement: The nature of racial segregation in everyday urban mobility. Urban Studies. https://doi.org/10.1177/0042098020978965

16.

Chetty

Hendren

Katz

(2016) The effects of exposure to better neighborhoods on children: New evidence from the moving to opportunity experiment. American Economic Review 106: 855–902.

17.

Coulton

Cook

Irwin

(2004) Aggregation issues in neighborhood research: A comparison of several levels of census geography and resident defined neighborhoods. In: APPAM Fall Research Conference. Atlanta, GA, November.

18.

Coulton

Korbin

Chan

, et al. (2001) Mapping residents’ perceptions of neighborhood boundaries: A methodological note. American Journal of Community Psychology 29(2): 371–383.

19.

Crowder

Pais

South

(2012) Neighborhood diversity, metropolitan constraints, and household migration. American Sociological Review 77(3): 325–353.

20.

Cutler

Glaeser

Vigdor

(2002) The rise and decline of the American Ghetto. Journal of Political Economy 107(3): 455–506.

21.

Cutler

Glaeser

Vigdor

(2008) When are ghettos bad? Lessons from immigrant segregation in the United States. Journal of Urban Economics 63(3): 759–774.

22.

Deville

Linard

Martin

, et al. (2014) Dynamic population mapping using mobile phone data. Proceedings of the National Academy of Sciences of the United States of America 111(45): 15,888–15,893.

23.

Ester

Kriegel

H-P

Sander

, et al. (1996) A density-based algorithm for discovering clusters in large spatial databases with noise. In: Simoudis

Han

Fayyad

(eds) Proceedings of the 2nd International Conference on Knowledge Discovery and Data Mining. 2–4 August 1996, Portland, OR. Palo Alto, CA: The Association for the Advancement of Artificial Intelligence, pp. 226–231.

24.

Fraser

Chaskin

Bazuin

(2013) Making mixed-income neighborhoods work for low-income households. Cityscape 15: 83–100.

25.

Freeman

(2009) Neighbourhood diversity, metropolitan segregation and gentrification: What are the links in the US? Urban Studies 46(10): 2079–2101.

26.

Galster

(2002) An economic efficiency analysis of deconcentrating poverty populations. Journal of Housing Economics 11(4): 303–329.

27.

Galster

Andersson

Musterd

(2010) Who is affected by neighbourhood income mix? Gender, age, family, employment and income differences. Urban Studies 47(14): 2915–2944.

28.

Galster

Andersson

Musterd

, et al. (2008) Does neighborhood income mix affect earnings of adults? New evidence from Sweden. Journal of Urban Economics 63(3): 858–870.

29.

Galster

(2012) The mechanism(s) of neighbourhood effects: Theory, evidence, and policy implications. In: van Ham

Manley

Bailey

, et al. (eds) Neighbourhood Effects Research: New Perspectives. Dordrecht: Springer, pp. 23–56.

30.

Galster

Friedrichs

(2015) The dialectic of neighborhood social mix: Editors’ introduction to the special issue. Housing Studies 30(2): 175–191.

31.

Galster

Killen

(1995) The geography of metropolitan opportunity: A reconnaissance and conceptual framework. Housing Policy Debate 6(1): 7–43.

32.

Goetz

(2003) Clearing the Way: Deconcentrating the Poor in Urban America. Washington, DC: The Urban Institute.

33.

Golledge

Stimson

(1997) Spatial Behavior: A Geographic Perspective. New York, NY: Guilford Press.

34.

Horton

Reynolds

(1971) Effects of urban spatial structure on individual behavior. Economic Geography 47(1): 36–48.

35.

Iovan

Olteanu-Raimond

Couronné

, et al. (2013) Moving and calling: Mobile phone data quality measurements and spatiotemporal uncertainty in human mobility studies. In: Vandenbroucke

Bucher

Crompvoets

(eds) Lecture Notes in Geoinformation and Cartography. Cham: Springer, pp. 247–265.

36.

Jiang

Ferreira

González

(2017) Activity-based human mobility patterns inferred from mobile phone data: A case study of Singapore. IEEE Transactions on Big Data 3(2): 208–219.

37.

Jiang

Fiore

Yang

, et al. (2013) A review of urban computing for mobile phone traces: Current methods, challenges and opportunities. In: Proceedings of the 2nd ACM SIGKDD International Workshop on Urban Computing. Chicago, IL, 11 August. New York, NY: Association for Computing Machinery, pp. 1–9.

38.

Jones

Pebley

(2014) Redefining neighborhoods using common destinations: Social characteristics of activity spaces and home census tracts compared. Demography 51(3): 727–752.

39.

Kwan

M-P

(2009) From place-based to people-based exposure measures. Social Science & Medicine 69(9): 1311–1313.

40.

Kwan

M-P

(2013) Beyond space (as we knew it): Toward temporally integrated geographies of segregation, health, and accessibility. Annals of the Association of American Geographers 103(5): 1078–1086.

41.

Lee

Lin

(2017) Natural amenities, neighbourhood dynamics, and persistence in the spatial distribution of income. The Review of Economic Studies 85(1): 663–694.

42.

Lieberson

(1981) An asymmetrical approach to segregation. In: Peach

Robinson

Smith

(eds) Ethnic Segregation in Cities. London; Canberra: Croom Helm, pp. 61–82.

43.

Liu

, et al. (2018) Mapping hourly dynamics of urban population using trajectories reconstructed from mobile phone records. Transactions in GIS 22(2): 494–513.

44.

Logan

Zhang

(2010) Global neighborhoods: New pathways to diversity and separation. American Journal of Sociology 115(4): 1069–1109.

45.

Lynch

(1960) The Image of the City. Vol. 11. Cambridge, MA: MIT Press.

46.

Manley

van Ham

, et al. (2011) Social mixing as a cure for negative neighbourhood effects. In: Bridge

Butler

Lees

(eds) Mixed Communities: Gentrification by Stealth?Bristol: Policy Press, pp. 151–168.

47.

Massey

(2016) Residential segregation is the linchpin of racial stratification. City & Community 15(1): 4–7.

48.

Massey

Denton

(1989) Hypersegregation in U.S. metropolitan areas: Black and Hispanic segregation along five dimensions. Demography 26(3): 373–391.

49.

Massey

Tannen

(2015) A research note on trends in black hypersegregation. Demography 52(3): 1025–1034.

50.

Matthews

(2011) Spatial polygamy and the heterogeneity of place: Studying people and place via egocentric methods. In: Burton

Matthews

Leung

, et al. (eds) Communities, Neighborhoods, and Health. New York, NY: Springer, pp. 35–55.

51.

Ostendorf

Musterd

Vos

(2001) Social mix and the neighbourhood effect. Policy ambitions and empirical evidence. Housing Studies 16(3): 371–380.

52.

Palmer

JRB

Espenshade

Bartumeus

, et al. (2013) New approaches to human mobility: Using mobile phones for demographic research. Demography 50(3): 1105–1128.

53.

Phillips

Levy

Sampson

, et al. (2019) The social integration of American cities: Network measures of connectedness based on everyday mobility across neighborhoods. Sociological Methods and Research, DOI: 10.1177/0049124119852386.

54.

Putnam

(2007) E Pluribus Unum: Diversity and community in the twenty-first century. The 2006 Johan Skytte Prize Lecture. Scandinavian Political Studies 30(2): 137–174.

55.

Reades

Calabrese

Sevtsuk

, et al. (2007) Cellular census: Explorations in urban data collection. IEEE Pervasive Computing 6(3): 30–38.

56.

Robson

Butler

(2001) Coming to terms with London: Middle-class communities in a global city. International Journal of Urban and Regional Research 25(1): 70–86.

57.

Sampson

(2011) Great American City: Chicago and the Enduring Neighborhood Effect. Chicago, IL: The University of Chicago Press.

58.

Schwartz

(2014) Housing Policy in the United States. New York, NY: Routledge.

59.

Smith

(2018) Analysis of Historical Traffic Speeds in Chicago. Chicago, IL: Chaddick Institute for Metropolitan Development, p. 4.

60.

Vallée

Cadot

Grillo

, et al. (2010) The combined effects of activity space and neighbourhood of residence on participation in preventive health-care activities: The case of cervical screening in the Paris metropolitan area (France). Health & Place 16(5): 838–852.

61.

Wang

Phillips

Small

, et al. (2018) Urban mobility and neighborhood isolation in America’s 50 largest cities. Proceedings of the National Academy of Sciences of the United States of America 115(30): 7735–7740.

62.

Wilson

(1987) The Truly Disadvantaged: The Inner City, The Underclass, and Public Policy. Chicago, IL: University of Chicago Press.

63.

Wong

Shaw

S-L

(2011) Measuring segregation: An activity space approach. Journal of Geographical Systems 13(2): 127–145.

64.

Xiao

Xiong

(2015) Protecting locations with differential privacy under temporal correlations. In: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security. Denver, CO, 12 October. New York, NY: Association for Computing Machinery, pp. 1298–1309.

65.

Yao

Liu

, et al. (2017) Mapping fine-scale population distributions at the building level by integrating multisource geospatial big data. International Journal of Geographical Information Science 31(6): 1220–1244.

66.

Yinger

(1986) Measuring racial discrimination with fair housing audits: Caught in the act. The American Economic Review 76(5): 881–893.

67.

York Cornwell

Cagney

(2017) Aging in activity space: Results from smartphone-based GPS-tracking of urban seniors. The Journals of Gerontology: Series B 72(5): 864–875.

68.

Yuan

Raubal

(2012) Extracting dynamic urban mobility patterns from mobile phone data. In: Xiao

Kwan

Goodchild

, et al. (eds) Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7478 LNCS. Berlin; Heidelberg: Springer, pp. 354–367.