The Role of Demographic Data Bias in the Under-Provision of Retail: A Quasi-Experimental Case Study of Low Food Access Neighborhoods in Twenty-Five Southern Metros

Abstract

Municipal policy attempting to remediate low food access neighborhoods tends to focus on improving demand conditions in these neighborhoods. We investigate the role of two fundamental measures of food demand (population and income) and the biases inherent in these data in creating low food access neighborhoods. Population and income data were collected for a 1-mile radius surrounding seventy-one grocery desert sites in Southern U.S. metro areas, those results were compared to thirty-eight low-income, non-desert sites in the same metros. No significant difference between the demand characteristics of desert and non-desert sites was found in our sample—suggesting that policy may need to be refocused on issues other than demand metrics. In addition, we detected significant demand underestimation bias from one source commonly used by grocery stores. Given these findings we believe that parcel level characteristics such as visibility, accessibility, and buildability may play a larger role in remediating low food access than addressing demand conditions.

Keywords

food access site selection spatial demographic data retail underprovision

Introduction

Social scientists and public health professionals have written volumes on the availability of fresh food in urban neighborhoods. This scholarship has focused on the issue of food availability from the perspectives of public health (Shannon 2014), race (Raja, Ma and Yadav 2008), and planning (Rose and Richards 2004). Despite the nuanced and varied view that academics have taken to analyze the causality of low food access, policy makers tend to fall back on the assumption that these food deserts¹ are primarily a result of weak demand. The discontinuity between academics and policy makers is exemplified by the current mayor of Charlotte, who shortly after her election in 2017, spoke about the potential role of the city in grocery desert remediation. She stated that it was the city's job to “help improve those metrics [of neighborhood demographics]” rather than attempt to encourage grocery operators to locate in underserved neighborhoods in the absence of attractive demographics (Peralta 2018). This well-intentioned perspective, which can also be seen in other cities in the American South (Hong 2019; Murphy 2019), as well as in academics (Myers and Caruso 2016; Hamidi 2020), assumes that grocery sellers have accurate information about neighborhood demand conditions and that neighborhood demographics are a first-order concern to these site selection professionals.

Following a recent study which found commercially produced demographic data to be less accurate in low-income areas (Graves and Gerney 2018), this note seeks to evaluate the impact of these biases on grocery store site selection in low-income neighborhoods. Specifically, we test the importance of two traditional metrics, population, and income, in grocery store site selection after controlling for race and urban form at the neighborhood scale using a quasi-experimental approach. By comparing the demographic characteristics of thirty-eight existing grocery store sites in low-income neighborhoods to the same characteristics of potential store sites in seventy-one locations in U.S. Department of Agriculture (USDA)-designated low food access neighborhoods (Figure 1), we found no evidence that demographics differ between food desert and non-desert areas in our sample. More significantly, we found little consistency in reported site characteristics between data vendors—this finding indicates that some sites may be rejected by grocery retailers due to their choice of demographic data provider. Given that demographic data vendors have been shown to underestimate population and income at the tract level in low-income neighborhoods (Graves and Gerney 2018) our findings indicate that some low food access areas may be a product of vendor data choice on the part of retailers. The purpose of this exercise is to provide local policy makers with more information about how site characteristics are assessed in chain-store location in low-income areas and to offer a novel explanation for how some food desert sites are produced.

Figure 1.

Twenty-five metro areas included in study. Seventy-one desert areas and thirty-eight non-desert areas were analyzed across the twenty-five metro areas.

Background

Academic research on grocery deserts has evaluated a broad range of causal mechanisms. These causes range from place-based (e.g., lack of suitable sites), demographic (e.g., income and race) to global (e.g., the consolidation of the retail industry). Since aggregate U.S. data, such as the USDA Food Atlas, shows a strong inverse relationship between income and the probability of low food access (Pothukuchi 2005), municipal remediation policy frequently focuses on strategies to subsidize new store operation in low-access neighborhoods to supplement local demand that is thought to be insufficient. The retail location process traditionally relies on measures of population and income as proxies for demand (Ghosh and McLafferty 1987; Hamidi 2020), and low population density and income is commonly cited by store chains as the primary cause of low food access (Zhang and Ghosh 2015).

This focus on demand conditions can lead scholars to conceptualize grocery deserts as “anomalies in an otherwise functional food system” (Shannon 2014, 256)—a perspective which may lead scholars to assume that correcting demand conditions is sufficient to remediate problems of food access. This perspective distracts policy makers away from exploring the underlying structural mechanisms (e.g., zoning, urban form, corporate consolidation, and marketing strategies) that create inefficiencies in the food distribution system (Gittell and Thompson 1999). Exacerbating the problem of food access has been the recent expansion of Dollar Stores into low-income urban neighborhoods. These stores sell a limited range of foods (generally no produce or unprocessed proteins) but create enough competition to discourage grocery stores from locating nearby (Shannon et al. 2018).

Site Selection and Food Access

Site selection professionals tend to consider the location decision process to be an exercise in rational optimization—businesses are thought to only build stores in locations that maximize revenue while minimizing competition (Sadler 2016). There is, however, a growing awareness among scholars of systemic biases in the site selection process which discourage store locations in low-income communities, even when demand measures indicate the site would be profitable (Shannon 2014). One of these biases is that the relatively small parcels available in urban areas are poorly suited for the large size of modern grocery stores (Ledoux and Vojnovic 2013). A second bias against low-income areas is the gravity created by large suburban stores which syphon the most affluent consumers away from smaller intown retailers (Guy, Clarke and Eyre 2004). Finally, the nature of demographic data used by grocery retailers for site selection appears to bias retailers against low-income urban neighborhoods as well. The commercially produced data has been shown to consistently underestimate income and population in low-income areas (Graves and Gerney 2018). Difficulty in measuring demand in low-income areas has led to the suggestion by Shannon et al. (2018) that scholars consider more direct measures of demand such as federal Supplemental Nutritional Assistance Program (SNAP) payment data when examining low-access neighborhoods.

Despite obstacles impeding grocery store expansion in low-income areas, there are chains that target low-income neighborhoods. At the time of this analysis, these chains include Food Lion, Food City, Sav-A-Lot, and Harvey's in the Southeastern United States. Location professionals working in these companies do not consider neighborhood incomes to be a first-order demand metric since food consumption is considered to be relatively inelastic (Norberg 2017; Rincón and Tiwari 2020). Unfortunately, low profit-margins create a high degree of cost-sensitivity for these chains, making sites that impose high development costs (e.g., oddly shaped parcels, zoning restrictions, or access difficulties) cost prohibitive (Ghosh-Dastidar et al. 2017).

Errors in Vendor-Provided Demographic Data

While the site selection process for retailers such as grocery stores relies on a multiscaler array of variables ranging from information about the parcel to the metro area, demand is fundamentally tied to the number of consumers and their income near the store site. Other indicators of store demand and potential operating expenses will likely include the change in population and income, the number of high(er) income households in an area, employment, racial composition, crime, and other variables (e.g., change in home value) that may indicate neighborhood change is underway (Rummo et al. 2017; Ge et al. 2019). Modern corporate-scale store site selection relies on the comparative analysis of large volumes of demographic variables. Unfortunately, there is little critical examination of this input data (Graves and Gerney 2018). Municipal officials often assume that grocery chains select sites objectively and that the input data used in their decision process are accurate. By overlooking possible biases in these data, policy makers may be proposing ineffective or inefficient solutions to the problem of food access when they assume site selection is a rational process.

Grocery retailers’ need for real-time data in their site decision process requires them to purchase spatial demographic data from vendors. These vendors model current year conditions from recent ACS data augmented with proxy variables (such as the USPS Master Address file) to estimate population change. Corporate data users know little about the accuracy of this vendor-produced demographic data. The details of these estimation models are proprietary (see Experian (2012), ESRI (2014), Synergos (2016), and Scan/US (2016) for their brief discussions of estimation methodologies) and the cost of conducting large-scale comparisons of vendor data is prohibitive. The foundation of these estimations, the ACS, is considered by most to be the highest quality demographic information available outside of the decennial Census (Meyer, Mok and Sullivan 2015). However, it is important to keep in mind that the ACS is produced from relatively small survey samples, which introduces errors at the subcounty scale, particularly in low-income areas (Bazuin and Fraser 2013). These errors are then propagated into all vendor estimates.

The lack of critical examination of these data is partially explained by corporations relying on vendor-provided data of current year conditions rather than the more accurate (but dated) Census data that scholars use when evaluating demographic phenomena. Unfortunately, subcounty Census data is not typically available until approximately two years after data collection, a lag that makes the data unworkable for store location professionals. Current year demographic data from vendors such as ESRI, Experian, or PopStats are derived using algorithms that modify older ACS data using proxy data such as the USPS Master Address File—a process which the USPS suggests is inappropriate for small area forecasts, particularly in high-density and low-income areas (Swanson and McKibben 2010). Graves and Gerney (2018) compared vendor estimates to ACS surveys of population and income (which are released two years after the vendor data are published for a given year). This study found that vendor estimates of tract-level population and income had higher error rates (compared to ACS) in high-density and low-income areas. Vendor population estimates deviated an average of 27% against ACS estimates and income deviated by an average of 40% against ACS figures in the sampled urban tracts. It was also found that vendors were more likely to overestimate population in low-income areas and underestimate incomes—a pattern that may cause grocery chains to see artificially low estimates of buying power (BP).

While corporate location analysts rely on a wide array of variables in their decision-making process, current-year, vendor-produced demographic data is, from necessity, modeled. The same biases that create lower population and income estimates in low-income neighborhoods are likely to be present in the estimation process for other current year variables. In short, current year estimation process is expected to create similar inconsistencies in other demographic variables used in the corporate site selection process.

Study Design

We first set out to test the expectation that grocery site selection in low-income neighborhoods is a rational and efficient process in terms of optimizing two foundational elements of grocery demand, proximate population, and income. As discussed above, municipal officials commonly assume that the problem of low food access is driven by these demand characteristics. In addition, we assumed that grocery desert locations will have worse demand characteristics than nearby non-desert sites. Specifically, we expect that our full sample of 109 sites will reveal that grocery desert sites have lower population and incomes within a 1-mile radius than non-desert sites. While this test is an oversimplification of the actual store location process, it reflects the perspective that municipal officials bring to the food desert remediation process.

Secondarily we evaluate the consistency of the spatial demographic data provided by the four data vendors. Our analysis focused on population and income since these two variables are foundational elements of demand and the two variables are used to estimate more specific measures (such as food consumed at home) provided by data vendors. We expect that vendors will provide data that is virtually identical to other vendors and any variation between vendors will be random. In short, we expect the site selection decisions of grocery retailers to be consistent regardless of data vendor used. If this is not the case, then data-vendor choice, rather than actual demand conditions, may be a contributing element to the uneven landscape of food access.

To test these expectations, we created a quasi-experimental study design which allows the comparison of a group of low-access neighborhoods, to a similar set of neighborhoods with a full-line grocery store. This approach allows for differences in competition, race, population and income growth, and urban form to be controlled for between the subject and control sites. This counterfactual technique borrows from Isserman and Merrifield’s (1987) study of regional growth and has been used to evaluate the impact of opportunity zones on employment growth (Arefeva et al. 2020) as well as housing development and neighborhood stability (Delmelle et al. 2017).

Study Area

We examined the demographic characteristics of low food access neighborhoods in twenty-five metro areas across ten Southern states² (Figure 1). The Southern United States (excluding Louisiana and South Florida) was selected as our study area due to the region's homogeneity in terms of auto dependence, urban form, racial composition, food culture, and income inequality. This regional framework was selected to control for differences in mobility, food culture, and density may impact the commercial site selection process. Our study area consists of metropolitan areas which ranged in population from just over 200,000 (Augusta, GA) to nearly 6 million (Atlanta).

Neighborhoods were selected using descriptive statistics and visual analysis of areal imagery to insure consistency in neighborhood type, including racial composition. This was verified by testing for differences in the mean values of African American population as a percent of total population and income growth in our two groups of neighborhoods. No significant differences in percentage of African American population, rates of population growth, or income change were detected between control and subject groups (at p > .05). In addition, fieldwork supported these findings of no significant difference in twelve of our twenty-five metro areas.

Selection of Low Food Access Areas

Seventy-one census tracts that are considered low-income, low-access (e.g., grocery desert) tracts were identified using the USDA Food Access Research Atlas (Economic Research Service 2017). The Atlas defines low-income, low-access areas as census tracts with either:

Poverty rates in excess of 20%

Median family incomes of less than 80% of the metropolitan area median

Plus:

At least 500 residents (or 33% of the census tract's population) who are more than ½ mile from the nearest supermarket

Within these low-income, low-access areas, we identified the site (a building) closest to the tract's centroid that appeared to have the characteristics (size, zoning, visibility, and access) necessary to be the home of a grocery store. These sites were used as the center of a 1-mile radius trade area for study. These sites were identified using the following criteria:

Air photo views indicated sites were zoned for commercial activity (e.g., developed land, distinct from residential areas, sufficient lot and building size, parking availability, access to adjacent roadway)

Sites were adjacent to a high-traffic thoroughfare

Sites were not immediately adjacent to single family residential

Sites contained active commercial business wherever possible (Table 1)

Table 1.

Current Activity at the Feasible Sites Used to Analyze the Demographic Characteristics of the Grocery Desert Areas.

Sites	Current use
17	Convenience stores (including gas stations)
23	Dollar stores
17	Restaurants (primarily fast food)
14	Various retail (including banks, shopping centers, and one vacant commercial space)

This process yielded seventy-one sites within the desert tracts (one per low-access tract) that appear to have the requisite zoning, visibility, and access characteristics needed to house a viable grocery store. Twelve tracts were removed from our initial eighty-three site set due to the absence of a viable grocery site per the process outlined above. Table 1 summarizes the business types of stores occupying these seventy-one parcels at the time the air photos used for our site selection were taken. Based on imagery, only two of the storefronts used as potential grocery sites appeared to be vacant at the time of data collection.

Selection of Control Areas

An additional thirty-eight low-income neighborhoods were identified in our metro areas that have an operating full line grocery store. These locations served as our non-desert control group set—sites in low-income communities which have demonstrated that a grocery store can be operated profitably (profitability is assumed from the existence of the active grocery store). The control group will be used as a baseline demand comparison for the low food access sites.

All but three of our low-income neighborhood grocery stores were operated by one of the following chains: Sav-A-Lot, Harvey's, Piggly Wiggly, Food Lion or Food City. These stores were all located in tracts where per capita income was less than $25,000 and were urban as defined by proximity to the city center as well as relatively high residential densities.

Demographic Data Collection

Data on population and per capita income were collected for the 1-mile radius surrounding each desert site (the identification of these addresses is described above) as well as each of our control store sites. The 1-mile radius was chosen based on the suggestions from active site selection professionals (Norberg 2017) as well as previous academic studies of food access (Cooksey-Stowers, Schwartz and Brownell 2017). We acknowledge this trade area size to be arbitrary as it does not reflect the nuances of the site selection process of every grocery vendor; however, we believe this distance is a reasonable proxy for typical urban trade area size in the Southern United States and the 1-mile radius is used here simply as an analytical device to facilitate comparisons. Previous research on the accuracy of vendor data indicates that our findings should be robust to larger trade areas. Data were retrieved in January 2018 from PopStats, Experian, ESRI, Scan/US. The default settings in each vendor's software were used when retrieving the data.

Study Limitations

The nature of our quasi-experimental strategy has, by necessity, limited the range of environments where we can evaluate the role of data vendor inconsistency in the formation of food deserts. Our sample limits our discussion to metro areas in the Southern United states. Our sample was further restricted to high-density neighborhoods (relative to other Southern U.S. neighborhoods) which have a high percentage of African American population. As such our findings may not be generalizable to neighborhoods with a different racial makeup, rural areas, or food desert areas in other regions.

Findings

Expectation: Data vendors will provide consistent demographic estimates for each site.

The data produced by vendors are intended to be accurate estimates of current, on the ground conditions—they all are intended to measure the same objective reality. Unfortunately, the vendor estimates of the same locations will substantially vary. This inconsistency is most easily seen at the site scale. Table 2 shows reported demographic characteristics for one desert and one non-desert site in Durham, NC. These data reveal a common level of variation in trade area population and income estimates at the site scale. The two sites in Table 2 show individual population estimates to vary by 20–30% between vendors and income estimates can vary as high as 54% between vendors (a $6,765 difference in per capita income estimated for the same site by different vendors). While data on these two sites reveals little about larger trends, they are indicative of the extent of estimation-model inconsistency between vendors. These data also allow us to see the effect of a grocery retailer's data choice on their perception of site quality.

Table 2.

Variation in Reported Site Characteristics Between Data Vendors for Two Sites in Durham, NC. Data collected from Each Data Provider for a 1-mile Radius Around Each Address.

Non-Desert 812 Liberty St			Desert: 1805 Miami Blvd
	Pop.	Per capita inc.		Pop.	Per capita inc.
PopStats	17,077	$19,230	PopStats	4,470	$35,423
Experian	14,023	$12,465	Experian	4,057	$30,401
ESRI	14,928	$13,189	ESRI	3,419	$31,567
Scan/US	15,696	$15,582	Scan/US	3,882	$37,267
Max difference	21.8%	54.3%	Max difference	30.7%	22.6%

To create a more comprehensive visualization of site-level variation in vendor estimates, Figure 2 shows vendor estimates of trade area BP. The BP is simply the product of the trade area's population and per capita income, the measure is used to assist in our visualization of variation in the market potential across an array of trade areas—we are not suggesting that BP is a commonly used metric for site selection. Each vertical column in Figure 2 represents an individual site while each of the four vertical dots display the BP from each vendor's estimates. The BP figures for our 1-mile radial trade areas range from $50 to $400 million dollars. Using the first column in Figure 2 as an example, PopStats estimates the highest BP ($158 million) in the trade area, while ESRI estimates the lowest market potential for the same area ($107 million).

Figure 2.

Estimated buying power (BP) calculated from each vendor's estimates. Sites are ordered from low population density to high population density. Desert sites are on the left of the graph; non-desert sites are shown in the shaded area at the right of the graph. The 50% of sites with the lowest variation are not shown.

This $51 million dollar difference in BP in the example above is substantial in the context of a typical, mid-sized, grocery store. The USDA Food Expenditure Survey estimates that 4.5% of an American's income is spent on food at home. Given the 4.5% figure, the $51 million disparity in ESRI and PopStats BP estimates translates into a $2.3 million variation in projected grocery sales within the trade area. According to the USDA (2019) this is more than 15% of a typical mid-sized grocer's annual sales, a projected revenue variation large enough to influence the profitability of an individual store. The mean variation in BPI estimates for all our desert sites was $34 million (11% of typical store sales). Our analysis indicates that variations between vendor estimates are large enough in 89% of our desert sites that a change in data vendor would skew store sales projections by more than 10%. This level of variation is sufficient for data vendor choice to discourage store location in the vast majority of our desert sites.

Expectation: Store sites in low food-access neighborhoods will have worse demand characteristics than nearby non-desert sites.

We set out to test the expectation of municipal officials, such as Charlotte's mayor, that grocery deserts are commonly created by weak neighborhood-scale demand conditions. Our initial test of this expectation gathered each vendor's population estimates and averaged them across our seventy-one desert sites and separately for the thirty-eight non-deserts sites. The same calculation was made for per capita income estimates from the four vendors (Table 3). Full-sample means were used to smooth volatility produced by single, outlier, sites. In theory these means should be equal for each of our four vendors since each is attempting to estimate the same objective reality.

Table 3.

Comparisons of Aggregate Population and Per Capita Income Estimates by Data Vendor. The Frequency of Estimates That Fall Below the Four-Vendor Mean are Also Shown as a Means of Illustrating the Significance of Over and Underestimation Bias.

Population	PopStats		Experian		Esri		Scan/US
Population	Desert	Non-desert	Desert	Non-desert	Desert	Non-desert	Desert	Non-desert
Mean	8,927	9,047	9,212	9,297	9,264	9,218	9,310	9,416
Percentage of observations below 3 vendor mean	68%	63%	37%	37%	35%	55%	44%	42%
No. of significant difference from the remaining three vendor mean*	3	1	1	0	1	0	1	1
Per capita income	PopStats		Experian		Esri		Scan/US
	Desert	Non-desert	Desert	Non-desert	Desert	Non-desert	Desert	Non-desert
Mean	$17,446*	$15,544	$17,901*	$15,617	$17,697*	$15,893	$18,597*	$16,235
Percentage of observations below 3 vendor mean	70%	71%	48%	61%	55%	55%	31%	26%
No. of significant difference from the remaining three vendor mean*	2	1	2	0	1	0	3	1

*Significantly different at .05.

In terms of population, we found that, in aggregate, each vendor reported a higher mean population for non-desert sites than desert sites (with the exception of ESRI which reported marginally higher populations for desert sites). While the population estimations were as we expected, none of the differences between desert and non-desert sites between vendors were statistically significant in our sample. This lack of variation in this coarse analysis indicates that the reported population, on its own, was not related to the existence of a grocery deserts in our sample. Our findings for the income metric were contrary to the expectations of policy makers. Every vendor reported higher mean incomes for desert trade areas than the non-desert trade areas. We found these differences between desert and non-desert sites to be significant at 95% confidence level. In the aggregate, finding that mean incomes are higher in desert sites in our sample indicates that income, on its own, was not related to the existence of a grocery desert within our sample. Together these comparisons suggest that the primary demographic characteristics of a site have relatively little influence on the neighborhood-scale availability of a grocery store. The absence of a significant relationship between demand conditions and store location may be explained by a variety of nondemographic factors such as the presence of a legacy store, zoning, parcel conditions, traffic counts, or nearby employment.

Given the amount of noise detected in the vendor estimates it is important to examine the variation of reported demographic data between vendors. In terms of population, PopStats reported the lowest numeric means (for both desert and non-desert sites) while Scan/US reported the highest—a trend which may be indicative of estimation biases between vendors. In terms of income, PopStats, Experian, and ESRI all provided income estimates of similar magnitude for both desert and non-desert sites while Scan/US provided income estimates that were substantially higher than other vendors. These findings reinforce the suggestion that data vendor choice may influence a grocery chain's views of a potential site's success. The trend of PopStats providing low estimates and Scan/US providing high estimates is consistent with the findings of Graves and Gerney (2018).

A second test of the magnitude of variation between vendors was conducted by comparing individual vendor mean values for population and income to the grouped mean of the other three vendors for desert and non-desert sites. The differences between individual vendors are statistically significant in nearly half of our comparisons; eight of twenty-four cross-vendor comparisons were significantly different for the population metric while ten of twenty-four were significantly different for the income metric (at p ≤ 0.05). These significant differences are frequent enough that vendor choice (rather than actual, on-the-ground conditions) appears to be impacting the geography of food access—our P-test results indicate that substantial, inconsistencies exist between the data reported by each of the four vendors.

Expectation: Inconsistencies in vendor data will be random.

The data above indicates that variations in these demographic data are large enough to discourage new store development in the majority of our sample sites. While it is often assumed that errors in estimated data will follow a stochastic pattern and thus cancel out in aggregate, Graves and Gerney (2018) found that these errors are nonnormal. We tested for normality in the inconsistencies between vendor estimates and the mean of the remaining three vendors in our sample using the Shapiro–Wilk test. Shapiro–Wilk indicated that none of the vendors produced data that had normally distributed around the remaining means. By tabulating the number of observations that fell below the remaining vendor's means, we assessed the tendency of each vendor to underestimate demand conditions (relative to the other vendors). This underestimation rate is also shown in Table 3. PopStats was found to most frequently underestimate the mean trade area demand conditions while Scan/US appears to consistently overestimate relative to other vendors. This underestimation bias in PopStats data is particularly troublesome for food-access in low-income neighborhoods since the majority of grocery chains surveyed by Graves and Gerney (2018) reported using PopStats for their location analytics in 2017. This will result in grocery chains perceiving worse demand conditions than actually exist. Since the errors in the vendor estimates are not randomly distributed, data vendor selection by grocery chains may result in neighborhoods being under-retailed.

Discussion

We examined estimation biases in two fundamental metrics of food demand: population and per capita income in grocery deserts, and low-income non-desert sites in an array of Southeastern United States metropolitan areas. Our sample of seventy-one grocery desert sites and thirty-eight low-income non-desert sites revealed no significant difference in population between the desert and non-desert sites, and mean per capita income was found to be significantly higher in desert sites. These findings were contrary to our expectations. In short, we found that differences in demand (as estimated by population and income) failed to explain the presence of grocery deserts within the low-income neighborhoods in our sample metro areas. While this finding is not meant to imply that demographic measures are irrelevant in the grocery location decision process, it does appear that the inelasticity of food demand may make demand conditions relatively less important in low BP neighborhoods. Instead we believe this finding indicates that site-specific factors such as zoning, access, visibility, and parcel shape may be more critical elements of store location decisions in this context.

In terms of variation between vendors we did detect the presence of bias rather than random estimation error. Our comparisons of each vendor's site values to the mean of the other three vendors revealed that these variations in vendor estimates were not random. Underestimation of BP was most prevalent in PopStats data while Scan/US consistently reported the highest demand figures. This finding indicates that, across our data set, grocery operators using PopStats data may be less likely to open a store in a low-income neighborhood than a grocery operator using other data providers. This finding of underestimation bias in PopStats data was consistent with other studies of vendor data bias in low-income neighborhoods.

While our analysis was limited to two fundamental measures of demand (population and income), the process of deriving other demand measures such as employment, consumer expenditure, and household size means it is likely the same estimation errors exist in those metrics. In other words, we expect the errors we detected in our sample sites’ population and income data to signpost additional errors in site selection processes which rely on secondary data. The magnitude of the estimation biases detected here was calculated using the BP for each trade area. The BP calculation indicated that differences in reported demographics between vendors could account for more than 10% of a store's projected income in the majority of our sampled sites.

When taken together, these findings suggest that marginal improvements in neighborhood demand conditions, the type of intervention that is often proposed by local officials, may be insufficient to remediate low food access conditions. In cases where local demand conditions are materially improved, our findings indicate that vendor-estimated demographic data may fail to detect those improvements.

Together our findings suggest that the location decision process for chain grocery stores is flawed and problems in the underlying data may contribute to the prevalence of low food access neighborhoods. We see three avenues for remediating these flaws. First, cities should emphasize buildability and other parcel-level characteristics of potential grocery sites. Promoting ready to build sites in low food access areas, similar to state-level industrial development policies, may encourage store construction. Second, cities may benefit from marketing potential grocery sites to retailers with ground-truthed demographic demand data, administrative data (such as SNAP) has shown promise in this context (Shannon et al. 2018). Finally, making grocery chains aware of the biases and flaws inherent in vendor provided demographic data will be necessary to develop a more efficient and equitable store location process. Unfortunately, this is a process that must occur in aggregate and is not within reach of an individual municipality.

More broadly, these findings indicate a need for policy makers and urban researchers to consider the role of flaws in the site selection process in the production of our built environment. Retailers are largely unaware of the biases inherent in their input data and, as retail continues to consolidate, these demand estimation errors are likely to compound retail inequities across the urban landscape. These inefficiencies clearly impact more than just grocery provision, many types of commercial and residential investment will be influenced by these data, and their associated biases, as well. At a minimum, academic and policy analysts need to be aware that commercial property developers are viewing the demographic landscape through a different lens than scholars.

Footnotes

Declaration of Conflicting Interests

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

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

William Graves

Notes

Author Biographies

William Graves, PhD, is an associate professor in the Department of Geography and Earth Sciences at UNC Charlotte. His research examines inefficiencies in site selection and how those inefficiencies influence urban form. Other lines of work include urban–rural economic interdependencies and the economic evolution of the American South.

Yi Zhang, MA, MBA, is an ArcGIS analyst at Esri, Redlands and recently completed a Geography MA from the Department of Geography and Earth Sciences at UNC Charlotte. She has worked for Yum! China, Domino's China, and Nike China and her research interests include retail network optimization and commercial site selection.

References

Arefeva

Alina

Davis

Morris

Ghent

Andra

Park

Minseon

. 2020. “Who Benefits from Place-Based Policies? Job Growth from Opportunity Zones.” Available at SSRN: https://ssrn.com/abstract=3645507.

Bazuin

Joshua

Fraser

James

. 2013. “How the ACS Gets It Wrong: The Story of the American Community Survey and a Small, Inner City Neighborhood.” Applied Geography 45: 292–302. https://doi.org/10.1016/j.apgeog.2013.08.013.

Cooksey-Stowers

Kristen

Schwartz

Marlene

Brownell

Kelly

. 2017. “Food Swamps Predict Obesity Rates Better Than Food Deserts in the United States.” International Journal of Environmental Research and Public Health 14 (11): 1366. https://doi.org/10.3390/ijerph14111366.

Delmelle

Elizabeth

Morrell

Elizabeth

Bengle

Tara

Howarth

Joe

Sorensen

Janni

. 2017. “The Effectiveness of Habitat for Humanity as a Neighborhood Stabilization Program: The Case of Charlotte, North Carolina.” Community Development 48 (4): 527–45. https://doi.org/10.1080/15575330.2017.1344717.

ERS (Economic Research Service), and USDA (U.S. Department of Agriculture). 2017. Food Access Research Atlas. Accessed November 10, 2017. https://www.ers.usda.gov/data-products/food-access-research-atlas/.

ESRI. 2014. Methodology Statement: 2014/2019 ESRI US Demographic Updates. White Paper. Accessed October 5, 2016. http://downloads.esri.com/esri_content_doc/dbl/us/J10268%20Methodology%20Statement_2014–2019%20Esri%20US%20Demographic% 20Updates.pdf.

Experian. 2012. Methodology Overview. Accessed October 5, 2016. http://alteryx.force.com/community/servlet/fileField?id=0BEE00000008OjG.

Dongdong

Luhui

Jiang

Guangjun

Xiaole

. 2019. “Intelligent Site Selection for Bricks-and-Mortar Stores.” Modern Supply Chain Research and Applications 1 (1): 88–102. https://doi.org/10.1108/MSCRA-03-2019-0010.

Ghosh-Dastidar

Madhumita

Hunter

Gerald

Collins

Rebecca

Zenk

Shannon

Cummins

Steven

Beckman

Robin

Dubowitz

Tamara

. 2017. “Does Opening a Supermarket in a Food Desert Change the Food Environment?” Health and Place 46: 249–56. https://doi.org/10.1016/j.healthplace.2017.06.002.

10.

Ghosh

Avijit

McLafferty

Sara

. 1987. Location Strategies for Retail and Service Firms. Washington, DC: Lexington Books.

11.

Gittell

Ross

Thompson

Phillip

. 1999. “Inner-City Business Development and Entrepreneurship: New Frontiers for Policy and Research.” In Urban Problems and Community Development, edited by Ferguson

Dickens

, 473–520. Washington, DC: Brookings Institution Press.

12.

Graves

William

Gerney

Brian

. 2018. “Evaluation of Commercial Demographic Data Accuracy: Testing for Input Bias in the Location Analysis Process.” Applied Geography 97: 152–60. https://doi.org/10.1016/j.apgeog.2018.06.009.

13.

Guy

Cliff

Clarke

Graham

Eyre

Heather

. 2004. “Food Retail Change and the Growth of Food Deserts: A Case Study of Cardiff.” International Journal of Retail and Distribution Management 32 (2): 72–88. https://doi.org/10.1108/09590550410521752.

14.

Hamidi

Shima

. 2020. “Urban Sprawl and the Emergence of Food Deserts in the USA.” Urban Studies 57 (8): 1660–75. https://doi.org/10.1177/0042098019841540.

15.

Hong

Christopher

. 2019. “City Hall Considers Program to Attract Grocery Stores to Northwest Jacksonville.” Florida Times-Union. May 14. https://www.jacksonville.com/news/20190514/city-hall-considers-program-to-attract-grocery-stores-to-northwest-jacksonville.

16.

Isserman

Andrew

Merrifield

John

. 1987. “Quasi-Experimental Control Group Methods for Regional Analysis: An Application to an Energy Boomtown and Growth Pole Theory.” Economic Geography 63 (1): 3–19. https://doi.org/10.2307/143847.

17.

Ledoux

Timothy

Vojnovic

Igor

. 2013. “Going Outside the Neighborhood: The Shopping Patterns and Adaptations of Disadvantaged Consumers Living in the Lower Eastside Neighborhoods of Detroit, Michigan.” Health and Place 19: 1–14. https://doi.org/10.1016/j.healthplace.2012.09.010.

18.

Meyer

Bruce

Mok

Wallace

Sullivan

James

. 2015. “Household Surveys in Crisis.” The Journal of Economic Perspectives 29 (4): 199–226. https://doi.org/10.1257/jep.29.4.199.

19.

Murphy

Ryan

. 2019. “A Year after Farm Fresh Closure, Norfolk Looks to Bring Piggly Wiggly to ‘Food Desert’ of Berkeley.” The Virginian-Pilot. May 14. https://www.pilotonline.com/government/local/article_18427dd0-766e-11e9-9b5f-7bbd55255117.html.

20.

Myers

Justin

Caruso

Christine

. 2016. “Towards a Public Food Infrastructure: Closing the Food Gap Through State-Run Grocery Stores.” Geoforum 72: 30–3. https://doi.org/10.1016/j.geoforum.2016.03.010.

21.

Norberg

Jeremiah

. 2017. Personal Conversation. November 28.

22.

Peralta

Katie

. 2018. “Charlotte Grocery Store Boom Continues to Skip Over Poor Areas.” Charlotte Observer. April 28. https://www.charlotteobserver.com/news/business/article215835705.html.

23.

Pothukuchi

Kameshwari

. 2005. “Attracting Supermarkets to Inner-City Neighborhoods: Economic Development Outside the Box.” Economic Development Quarterly 19 (3): 232–44. https://doi.org/10.1177/0891242404273517.

24.

Raja

Samina

Changxing

Yadav

Pavan

. 2008. “Beyond Food Deserts: Measuring and Mapping Racial Disparities in Neighborhood Food Environments.” Journal of Planning Education and Research 27 (4): 469–82. https://doi.org/10.1177/0739456X08317461.

25.

Rincón

Edward

Tiwari

Chetan

. 2020. “Demand Metric for Supermarket Site Selection: A Case Study.” Papers in Applied Geography 6 (1): 19–34. https://doi.org/10.1080/23754931.2020.1712555.

26.

Rose

Donald

Richards

Rickelle

. 2004. “Food Store Access and Household Fruit and Vegetable Use Among Participants in the US Food Stamp Program.” Public Health Nutrition 7 (8): 1081–8. https://doi.org/10.1079/PHN2004648.

27.

Rummo

Pasquale

Guilkey

David

Shu Wen

Popkin

Barry

Evenson

Kelly

Gordon-Larsen

Penny

. 2017. “Beyond Supermarkets: Food Outlet Location Selection in Four US Cities Over Time.” American Journal of Preventive Medicine 52 (3): 300–10. https://doi.org/10.1016/j.amepre.2016.08.042.

28.

Sadler

Richard

. 2016. “Integrating Expert Knowledge in a GIS to Optimize Siting Decisions for Small-Scale Healthy Food Retail Interventions.” International Journal of Health Geographics 15 (1): 19. https://doi.org/10.1186/s12942-016-0048-6.

29.

Scan/US. 2016. Detailed Estimates and Projections. Accessed October 5, 2016. http://www.scanus.com/specialty-detailed-demographic-data-products.htm.

30.

Shannon

Jerry

. 2014. “Food Deserts: Governing Obesity in the Neoliberal City.” Progress in Human Geography 38 (2): 248–66. https://doi.org/10.1177/0309132513484378.

31.

Shannon

Jerry

Bagwell-Adams

Grace

Shannon

Sarah

Lee

Jung Sun

Wei

Yangjiaxn

. 2018. “The Mobility of Food Retailers: How Proximity to SNAP Authorized Food Retailers Changed in Atlanta During the Great Recession.” Social Science and Medicine 209: 125–35.

32.

Swanson

David

McKibben

Jerome

. 2010. “New Directions in the Development of Population Estimates in the United States?” Population Research and Policy Review 29 (6): 797–818. https://doi.org/10.1007/s11113-009-9164-3.

33.

Synergos. 2016. STI: PopStats – Methodology. Accessed October 5, 2016. http://www.synergos-tech.com/ps_meth.htm.

34.

USDA (US Department of Agriculture), Economic Research Service. 2019. Food Expenditure Series. Accessed October 25, 2019. https://www.ers.usda.gov/data-products/food-expenditure-series.

35.

Zhang

Mengyao

Ghosh

Debarchana

. 2015. “Spatial Supermarket Redlining and Neighborhood Vulnerability: A Case Study of Hartford in Connecticut.” Transactions in GIS 20 (1): 79–100. https://doi.org/10.1111/tgis.12142.