An urban informatics approach to understanding residential mobility in Metro Chicago

Study area and data

Over the course of its history, Chicago has served as a national policy laboratory for low-income housing. The city has historically adopted several housing-related policy experiments such as the federal Moving to Opportunity for Fair Housing Demonstration Program and the HUD’s Moving to Work demonstration program. There is also a plethora of previous low-income housing studies in Chicago to compare with and complement the findings of this study, such as urban residential disparity (Chaskin et al., 2012), massive neighborhood types (Greenlee, 2019), accessibility and land use heterogeneity (Pan et al., 2020), and proactive policy measures (Chaskin et al., 2012). Lessons from Chicago can be used to inform cities worldwide in terms of the residential mobility of households from diverse socioeconomic backgrounds.

To measure longitudinal household locational characteristics, we used data from the InfoUSA consumer database. Researchers previously used InfoUSA data on business locations to measure local employment accessibility and labor market characteristics, particularly in studies focused on food deserts and other public health issues (Greenlee, 2019; Hoehner and Schootman, 2010; Moore and Diez Roux, 2006). The primary business use of the database includes targeting potential consumers’ mailing addresses in advertising campaigns. InfoUSA also maintains a consumer database that contains both historical and current household information by residential address. At the national level, InfoUSA’s consumer database contains information on more than 155 million households. These data are updated on a monthly basis, drawing household address information from the United States Postal Service, the National Change of Address database, the Locatable Address Coding database, and the Delivery Point Verification database. The data help to identify probable residential locational changes. Public records including tax assessor information and deed transfers are also used to assess household demographics.

In this study, each household is assigned a unique identification number that is maintained across the study years. This allows for the linking of yearly records into a longitudinal data set. Each yearly record reflects the household residential location in December of the year in question. Ten annual yearly “snapshots” (2006–2015) of residential locations are merged into one longitudinal database using the unique household identifier to organize records into longitudinal sequences. We used a 10-year average for three main reasons: (1) HUD median-income calculations have already taken inflation into account; (2) because of the economic crisis that occurred within the 10-year period, adjusted household income cut-off in Chicago does not have drastic changes; and (3) the income cut-off is only used to identify neighborhood level, not individual level; thus, the neighborhood income-level variations are limited in the 10-year period. We demonstrate the evidence supporting the three arguments in Supplemental material S1.

To determine household moves involving a change in neighborhood, we joined 2010 definition census tracts to our household address data. We flagged those households as moves who have a change in residential address that involves moving to a different census tract. We also created a separate data flag indicating a within census tract change in address.

Figure 1 shows the graphic representation of the origination and destination trajectories for households from low-, medium-, and high-income neighborhoods in Chicago over the study period. The OD comparison shows that mobility, to some extent, mitigates concentrated poverty for low income. However, the mitigation of concentrated poverty comes with a cost.Average moving distance of low-income households is 3.6 and 3.3 kilometers for extremely low-income households, which are both significantly higher than the average of 3.0 kilometers for medium and high-income households (and the differences are significant at a 95% confidence interval by pair-wise t-tests). Thus, low-income households need to move longer distances at higher costs to find better opportunities and neighborhood conditions. It is interesting and important to understand the comprehensive landscape of residential mobility in Chicago across the income spectrum to determine the driving forces of the results shown in Figure 1.

OPEN IN VIEWER

Figure 1.

Origination and destination trajectories of movers of low (upper left), medium (upper right), and high-income (lower left) household tracts in Chicago from 2006 to 2015.

Multi-dimensional spatial scan for OD data

Gao et al.’s (2018) multi-dimensional spatial scanning (MDSS) statistical approach for OD data has previously been used to compare morning and afternoon taxi trips in New York, and county-to county migration between age groups for the entire U.S. The approach successfully discovered trip identity and clusters beyond TAZ levels and migration groups beyond traditional identified borders (state or metropolitan).

This study adapts the MDSS method for three main reasons: (1) spatial scale: the multi-dimensional spatial scan can find clusters of individual household movements at a very fine spatial scale (100 meter scale); (2) OD data suitability: most spatial scan data are suitable for point or polygon use; the MDSS approach is specifically designed for OD data sets and uses a four-dimensional space to represent coordinates for each pair of OD locations; (3) model scalability: the method uses parallel processing techniques to analyze large amount of data (millions of households in this case) within a reasonable computational time.

The process of conducting MDSS is briefly described as follows (see Figure 2). First, every OD locational set is tagged with two pairs of X and Y coordinates. All the OD pairs are then randomly clustered—to test the null hypothesis that spatial data clusters are randomly organized. Multiple scanning windows are created with diameter i , over increment j , to final size s that cover all locations on the map ( i and s are parameters that will be explained in the next paragraph). In each four-dimensional scanning window, a Bernoulli likelihood ( l w ) is calculated to test the assumption that all OD pairs in the scanning window belong to the same clusters. A likelihood ( l 0 ) is also calculated to test the null hypothesis—whether the clusters in the scanning window are randomly assigned. Various sizes and locations of scanning windows are tested to find the maximum likelihood ratio ( l w / l 0 ) for our target assumption in this study—the scanning window features as the origin for one income group (among high, medium, or low) and the destination for another income group. As noted, when a scanning window features the origin of the higher income group and a destination within a lower income group, it is defined as a “flight” window. When a scanning window features the origin of a lower income group and the destination is higher, it is defined as a “gentrified” window.

OPEN IN VIEWER

Figure 2.

The flow of an MDSS process for MDSS applications in the case of Chicago residential mobility. OD: origination–destination.

Some key parameters for MDSS in this study include incremental interval j for scanning windows is 100 meters; maximum scanning window size s is set to 5000 meters; destination and origin scanning windows are non-overlapping; if a location is identified as both “gentrified” and “flight,” it is set to be “none.” For robustness, we tested different parameter settings for i and s and found no notable deviance from our study results.

Od movement type results

The Chicago region might be viewed as three concentric rings with the CBD as the center, second ring as the suburban extent, and third ring as the peri-urban or urban fringe. In the CBD, residential mobility is very diverse, and “flight” and “gentrification” usually occur in roughly the same places. Because of this, a pattern is not assigned using the spatial scan technique. In the Chicago region, the second ring includes first tier suburban areas—Blue Island, Oak Lawn, Cicero, Oak Park, Rosemount, Skokie, and Evanston, for example. These places have relatively large numbers of residents that move with typically obvious “flight” or “gentrification” identities. Our MDSS approach is most successful in identifying mobility patterns in these places. The third ring—the exurban fringe—is less densely populated and therefore has relatively little residential movement. In this zone, the MDSS does not recognize significant mobility patterns.

Figure 3 displays our “places” classification results using the MDSS spatial scan technique in the region. Our findings show that the south and west parts of Chicago feature the majority of low-income gentrification in Oak Park (west) and Hyde Park (south) and Hyde Park (south) by medium-income movers and Cicero (southwest) by high income. High- and medium-income flight is seen in Norridge (northwest) and the tip of Skokie (north). It is likely that the decline of the city’s manufacturing economy might play a role in this flight pattern. Rosemount and the O’Hare International Airport area (northwest) show a significant influx in medium- and high-income households that replace households from the lower income spectrum. These patterns might be related to previous findings by Yang et al. (2019) and McMillen (2003) that the O’Hare Airport region is an important employment growth hub and housing in this area is highly desirable for higher income households.

OPEN IN VIEWER

Figure 3.

The places classification results using the MDSS spatial scan technique in the seven-county region of Chicago.

One important observation is that although previous mobility policies have helped mitigate concentrated poverty in the CBD, neighborhoods with better socioeconomic conditions are more likely to gentrify with higher income households replacing low-income households. This type of movement is perpetuating the increasing divide between low- and high-income households. In Table 1, the category “low income gentrified by high income” shows the highest median household income, median house value, and the lowest vacancy rate, poverty rate, and housing cost burden. In comparison, neighborhoods with worse socioeconomic conditions tend to experience flights of high and medium households. The category “high-income flight from low-income” (Table 1) shows the lowest median house values, and the highest vacancy rate, poverty rate, housing cost burden, and unemployment rates. Table 1 further shows that “housing voucher assisted renter proportion” is higher in the high- and medium-income flight neighborhoods from low-income households’ inflow. This finding implies that low-income movement-triggered flights are associated with the existence of the assisted housing voucher program. Our results reveal some insights into the poverty decentralization campaign in Chicago (Chaskin et al., 2012). Some of the movement of low-income households may be the result of state-initiated programs aimed at mitigating concentrated poverty. On the other hand, if the movement exacerbates the exodus of medium- and high-income households, it can compromise goals of creating socioeconomically diverse neighborhoods (Chaskin et al., 2012).

OPEN IN VIEWER

Table 1.

Average socioeconomic statistics for different movement origin–destination (OD) types.

Movement type	Nonwhite proportion (%)		Median household income ($×1000)	Median house value ($×1000)	Vacancy rate (%)	Poverty rate (%)	Unemployment rate (%)	Housing Cost burden (% income spent on housing-related costs)	Housing voucher assistedrenter proportion (%)
Low income gentrified by high income	16	79		283	7	10	11	38	0
Low income gentrified by medium income	19	40		207	8	12	14	53	0
Medium income gentrified by high income	10	43		202	11	12	7	52	0
High-income flight from low income	81	28		157	23	42	24	55	19
Medium-income flight from low income	67	47		232	17	27	19	50	12
High-income flight from medium income	17	47		168	9	18	8	44	1

Source: Data from 2012 to 2016 American Community Survey—Table B07012.

Proposal for a PSS-assisted planning process: Residential mobility planning support system (RMPSS)

In this part, we propose the construction of a PSS interface based on the mobility big data and our analytics as a means of democratizing the MDSS methodology for informing movement decisions for low-income households—both by enabling them for more informed decisions and empowering them in collective policy-deliberation practices (such as public planning meetings) to argue for more equitable policy decisions. One central tenant of abating the cycle of poverty for low-income households is to help them move away from neighborhoods of concentrated poverty with assisted mobility programs of multiple levels of governance (McClure, 2006, 2008). To overcome limitations of assisted mobility programs in decentralizing poverty and potential difficulties for disadvantaged groups in accessing technology, PSSs have been found to be a useful way for policy-makers and planners to convene and facilitate public meetings and involve diverse stakeholders into participatory decision-making processes (Auerbach et al., 2018; Bailey et al., 2011; Pan and Deal, 2019).

We constructed a RMPSS to visualize the results of MDSS for Chicago and solicit feedback from local community members. One objective is to test data science and urban informatics techniques. The development of the RMPSS involves three key components: a cloud-based database, a replicable data model framework that translates data into meaningful outcomes for participation, and a front-end design that visualizes the model results for easy access (Figure 4). The interactive visual interface gives users easy access to model information and provides a forum for feedback and local knowledge input. It intends to bridge the gap between users and back-end models; helping extend stakeholder perception, especially for those without professional knowledge (Pelzer et al., 2013). A survey instrument is used to collect comments and as an information creator helping to provide feedback.

OPEN IN VIEWER

Figure 4.

The RMPSS framework from backend (database) to server to frontend (model) and user.

Figure 5 displays the visualized planning support interface and feedback portal (RMPSS). Further user surveys and the link to the website are recorded in Supplemental material S2. The interface features visual elements, explanation, and other graphics to convey the complex spatial data science methods used. The goal is to improve both the understandability of the analytical results and their use in public planning and policy formation. The map API allows users to zoom in and out to check specific locations and results, mark important locations (e.g., home, work), and overlay other maps (e.g. crime, census) to further explore model results and interactions. A dropdown for selecting other locations is also included and this allows for the expansion of the approach to other cities.

OPEN IN VIEWER

Figure 5.

The screenshot of the RMPSS visualized interface. RMPSS: residential mobility planning support system.

Policy implications

In this study, a model is used to translate large volumes of data into understandable and useful information for planning and policymaking. It is generally difficult for policy-makers and planners to find patterns or draw policy from the raw and tabular residential movement data, which has millions of entries. With the help of this model, the input raw data are processed via the model algorithm and users can easily obtain the visualized analytic results. In our approach, the raw data are made relevant for decision-making in three ways. First, it makes the data accessible by categorizing the OD of movement into understandable terms: gentrification and flight by the various income spectrums. Second, the OD model results are presented in a spatially explicit and highly visual format that can be easily evaluated by local users. It also complements previous research findings that characterize community statistics of the OD comparison of residential mobility (Chaskin et al., 2012; DeLuca et al., 2010; Keels, 2008). Third, the classification of movement destination types uses a resolution that goes beyond traditional administrative or census boundaries. Compared to previous studies of neighborhood mobility characteristics at the census tract level (Greenlee, 2019), our higher resolution scanning approach does identify different boundary patterns to clustering gentrification—flight types of mobility. Moreover, understanding the socioeconomics of the “outflow” and “inflow” of households at better spatial granularity can provide more accurate geolocation analytics and support location-targeted planning strategies, compared with the static and aggregated community-level characterizations (e.g. from census data). Our findings in Table 1 show the misalignment between socioeconomic characteristics of different mobility-type clusters to census tracts.

Previous studies of urban informatics focus on emerging big and open data sets to explaining the urban condition (Foth et al., 2009; Liu et al., 2015; Thakuriah et al., 2017). This study expands the idea to include urban “informatics”—the linkage of these data, to models, user-interfaces, and users (Bilandzic and Venable, 2011). The RMPSS takes advantage of big and spatially explicit data and makes use of advances in open-source computing technology by creating a replicable “model to interface process” through a fast-prototyping exercise of the whole system. We suggest that the urban informatics-based approach is not only useful in a research setting, but it is also important and integral to a data infrastructure with spatially explicit smart city frameworks. It can promote a socio-technical transition toward a new participatory design and planning paradigm.

An important limitation of this study is that we have not yet addressed several issues regarding advocating disadvantaged groups. First, identification of disadvantaged groups is important as a preliminary step for provision of information services. There are other studies that have focused on identification of disadvantaged groups in Chicago (Heller, 2014). Thus, there is a need to use expert knowledge, as well as consulting local knowledge, to identify disadvantaged groups in the locale. Furthermore, continuous local expertise consultation overcomes an identified issue of using statistical model for planning support—the inability to cope with ontological uncertainty (defined as surprises in the form of currently unknown future events—Derbyshire, 2019). To address this issue, it is necessary to use continued criticism from local expertise regarding the statistical model results and make corresponding updates.

Second, there is a need to better expose the technologies to disadvantaged groups, as it is a common issue that disadvantaged groups have less accessibility to and literacy of technologies (Wong et al., 2015). One potential solution to this issue is to apply PSSs in a policy-level rather than individual-level setting, by using the ability of PSSs to convene public stakeholder deliberations and actively involve disadvantaged groups (Jones, 2011; Klosterman, 1995; Yu et al., 2018). The third issue is related to the second issue as disadvantaged groups have limited access to the technology; thus, a web-based feedback solicitation system may not reflect the unbiased opinions from these groups. To address this issue, we confer that the identification of the community in question would help further forms of investigation. The potential approaches include participant observations that interact with targeted groups in everyday life as a means of collecting information from diverse groups (previously applied by sociologists such as Merton, 1945 and Jorgensen, 2015). The feedback solicitation mechanisms should help assess the usefulness and understandability of the interface and potential bottlenecks to use for disadvantaged groups (Vonk and Geertman, 2008; Vonk et al., 2005). Moreover, planners and modelers should be very careful in using the information. It should be limited to the stakeholder deliberation of policy-making processes with a special focus on empowering and informing positions of disadvantaged groups. A publication of the information with equal access to all—though seemingly an “equal” move—is likely to be exploited by, for example, real estate personnel working for the upper class to devise strategies in perpetuating inequity.

Another potential limitation relates to the question about whether income level as the only proxy for opportunity is biased or inadequate, although it is obvious that income correlates with other indicators such as educational attainment and crime rates. Furthermore, the system should be tailored to support for different housing types (single-family houses versus multifamily houses), different price ranges, and different neighborhoods (mixed housing types and prices, or homogenous housing units). These features are found to be related to the residential segregation issue and related findings would contribute to place-based policymaking by providing insights about how different housing features contribute to segregation (Owens, 2019).