The Implicit Price of Food Access in an Urban Area: Evidence from Milwaukee Property Markets

Study Area

Our study area is Milwaukee, Wisconsin, which is the largest city in the state. While few academic projects have studied Milwaukee in the context of food access, the city provides an interesting case study, in part due to its similarities to other urban, industrialized cities in the Midwest (McMillen 2001). These include cities such as Chicago and Detroit, each of which has been prominent in the discussion of food access in urban contexts (e.g., see Mari Gallagher Research and Consulting Group 2006 or Zenk et al. 2005). Much like those cities, Milwaukee has significant racial segregation.⁴ Neighborhoods to the north of the city contain a significant majority of AA households, while a large concentration of LA households are located in the south. These areas contrast with the eastern part of the city, positioned next to Lake Michigan, which consists of primarily young, white professionals. Outlying cities, such as Whitefish Bay, have white populations making up over 90% of the population, mostly consisting of older families. Appendix Figure A1 illustrates the distribution of AA and LA households in the city. This socioeconomic division suggests that the residential market in Milwaukee is distinct from those of the surrounding locales. Given this, we limit our study to include sales only within the city of Milwaukee proper.

Milwaukee also makes an appealing city to study due to its abundance of publicly available data. Estimating a hedonic model requires data characterizing home sale transactions, including the home’s structural features and neighborhood characteristics, including the location of relevant amenities and socioeconomic makeup. The city provides tabular and GIS data characterizing the landscape for free to the public, making the results here easily replicable.

Residential Transactions

The Milwaukee Assessor Office provides data detailing land sale transactions dating back to 2002. These transactions include single-family homes, as well as condominiums, vacant lots, and commercial properties. In addition to the sale year and price, the data provide information on basic structural characteristics of each property, including the number of bedrooms and bathrooms, lot and home size, and home style.⁵ For this study, we consider sales from the periods 2002–2006 and 2011–2015. The hedonic pricing model relies on the assumption of equilibrium in the market in question, an assumption likely violated during the recession years of 2007–2010. Appendix Figure A2 illustrates the number of home sales from 2002 to 2015, and Appendix Figure A3 shows the average real home price in 2002 dollars over the same period. Residential sales were fairly stable from 2002 to 2006 but fell sharply from 2007 to 2010, from 4,798 home sales to 815 in 2010. After 2010, home sales began to rise and did so steadily through the end of the study period. Real home prices experienced a similar fall during the recession years, from a high of $116,264 in 2006 to $100,411 in 2011. Unlike home sales, however, home prices continued to fall beyond this point. They stabilized in the final five years of the study period, averaging $95,253, likely corresponding to a stabilization in the housing market during this time. Together, this provides compelling evidence that the housing market was out of equilibrium during the Great Recession years. However, as a robustness check, we also run the same analysis with the full 14-year panel. After removing incomplete records, we also trim 2.5% of the records based on price from both tails of the distribution, and end up with 26,912 records used in this analysis. Table 1 reports summary statistics for these sales.

Land Use Data

To characterize the local amenities available at each home, including food access, we leverage Milwaukee’s Master Property Record (MPROP) and parcel shapefiles, which provide spatial and tabular data describing every parcel in Milwaukee dating back to 2001.⁶ Each parcel has a unique tax key contained in both the shapefiles and tabular record. This allows us to match each parcel’s structural characteristics with its location within the city. The same tax key is also present with the home sales transaction data, enabling us to match these transactions to their parcels as well.

The structural features found in the MPROP contain both physical characteristics of the buildings themselves, including additional variables not found in the home transactions data such as the presence of air conditioning, and land use characteristics. The MPROP contains two distinct land use codes. The first of these is a four-digit code based on the Standard Industrial Classification (SIC) code. Among these is a specific code for grocery stores, which we use as the basis for defining food access. A second land use variable separates the parcels into eight larger catch-all categories, such as commercial or residential.

Defining Food Access and Other Neighborhood Amenities

To define food access, we draw from the portion of the literature that defines food access using large, full-service grocery stores.⁷ While many stores, including convenience stores, offer food products, full-service grocery stores generally have lower prices (Kaufman et al. 1997) and a wider variety (Glanz et al. 2007), including selections of fresh meat and produce often not found in any single convenience store location. To proceed we must define large, full-service grocery stores and then derive a metric of food access. To begin, we sorted all of the locations coded as grocery stores present in the MPROP by their lot size for the year 2005.⁸ A clear break in grocery store size occurred at 10,000 sq ft, with the largest grocery store across the 10 years under that size reporting at 8,916 sq ft, and the next largest reporting at 10,406 sq ft.⁹ We then cross-checked the addresses provided by the MPROP for each of these stores with Google results, listings in the yellow pages, and store websites when available. While stores above this cutoff universally met the full-service criteria laid out by the District of Columbia, stores below this cutoff consisted solely of convenience stores (such as Seven-Eleven) and corner stores. As a result, we use 10,000 sq ft as the cutoff between “large” grocery stores included in this analysis, and stores excluded from the analysis.

In addition, we classify large department stores, namely, Super Walmarts and Targets, as grocery stores. While classified as department stores in the MPROP, they clearly fit the criteria of a full-service grocery store and fill that role for many households. For these, and a small number of grocery stores classified as “mixed used” parcels, we again used a combination of Google searches along with business listings to determine which addresses should be reclassified as grocery stores. In both of these cases, we manually changed their land use codes to match the one for grocery stores.

To define food access, we draw from the definition of a food desert provided by the USDA, which defines an urban food desert as a census tract where 33% of households do not have access to a quality food source within 1.0 mi of their residence.¹⁰ This definition captures the importance of “walkability” for food access in an urban area, where many individuals, particularly from low-income households, do not own a car. For this study, we define food access as the number of grocery stores found within half-mile rings up to 1.0 mi of a home. This definition refines the USDA definition to better capture the importance of walkability for local grocery stores. For a given household, it seems unlikely that it would value an additional grocery store a mile away the same as one only one-tenth of a mile away. However, given the limited number of grocery stores within a city, we argue that half-mile rings allow us to capture this effect while also keeping enough variation to identify the actual impact of a local food outlet. Recent papers by Pope and Pope (2015) and Walls, Kousky, and Chu (2015) have considered this heterogeneity in estimating the value of a nearby Wal-Mart and of landscape views, respectively. Figure 1 illustrates the location of grocery stores relative to single-family home transactions in 2005.¹¹

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Grocery Store Locations (black/red triangles) Relative to 2005 Home Sales

Excluding the recession years, the number of grocery stores varies from 52 to 56, suggesting a modest level of entry and exit from the market during our study period. We do not explicitly model this entry and exit for several reasons. First, in spite of entry and exit, the variation in stores from year to year is small. The number of outlets by year is illustrated in Appendix Figure A5. In line with the previous research on grocery stores and locational decisions, many of the observed entries occur in locations that were recently vacated by another store.¹² More pragmatically, the small number of entries and exists and the limited number of repeat sales observed in our data (less than 10% across 10 years) make a quasi-experimental approach using these dynamics infeasible.

One concern in identification is omitted variable bias. Omitted variable bias in the hedonic model arises from structural or neighborhood characteristics not included in the regression, but correlated with observable features. As retail locations, grocery stores typically appear in close proximity to other retail storefronts. The stores could act as either an amenity to households (employment opportunities, additional shopping options) or a disamenity (congestion), raising the concern for bias. To account for this, we include two measures of commercial land use, both at the property level: the percentage of land within a mile that is denoted as commercial use, and the distance to the nearest nongrocery commercial parcel.

As an additional commercial control, we include the number of bars within half-mile rings up to 1.0 mi of each home sale. Within a city, there is likely significant heterogeneity in the “type” of commercial lots in a given neighborhood. Areas with a high density of coffee shops may have a different impact on home prices than an area densely populated by fresh fish shops, though both would be represented the same way with the controls described above. The choice of bars as a more specific retail control has several advantages. Functionally, they have their own land use code in the MPROP, making them easy to establish. In addition, there is evidence to suggest that bars appear in high concentrations in urban areas, even in those with a lack of food outlets. Previous research has suggested that while food sources tend to diminish in communities with a high percentage of people of color, the number of bars does not (Moorland et al. 2002). In addition, Wisconsin in particular has a high density of bars, with previous studies suggesting that bars outnumber grocery stores by a factor of nearly three, as illustrated in Appendix Figure A4. That bars appear in dense commercial areas even in the absence of grocery stores make them an ideal additional control for commercial density. Finally, high concentrations of bars and liquor stores have been associated with negative amenities in previous research (see, e.g., Romley et al. 2007), allowing for the possibility of negative bias if not accounted for.

Generating Spatial Fixed Effects

Spatial heterogeneity and its impact on home prices presents another challenge to producing unbiased estimates in hedonic studies. Recent advances in the literature have used quasi-experimental models that exploit natural breaks in the data (see, e.g., Linden and Rockoff 2008 or Pope and Pope 2012 and 2015) as a preferred way to address omitted variable bias. However, such an approach is not always possible given their data requirements, or desirable due to their often highly localized contexts (Bajari et al. 2012). In this case, the use of spatial fixed effects has become standard to address unobserved local factors that may vary across space. Recent studies have gone beyond the use of spatial fixed effects in isolation, interacting them with time fixed effects or even other covariates (such as prices within a neighborhood from the previous year), in order to additionally control for variables that vary across time and space (Linn 2013; Kuminoff and Jarrah 2010; Anderson and West 2006).

In this study, given the relatively small sample size, as well as the low number of repeat transactions, which consist of less than 10% of our sample, a quasi-experiment approach is not feasible. Therefore, addressing the endogeneity of grocery store locations and other unobservable community features requires the use of spatial controls. Examples of unobserved features here may include the relative density of public transportation options, or broader employment opportunities represented by grocery stores and other retail locations.¹³ To address this, we derive a set of spatial fixed effects from data characterizing neighborhoods and high school districts in Milwaukee. Shapefiles characterizing high school districts were obtained from the University of Wisconsin–Milwaukee.¹⁴ For neighborhood descriptions, we rely on public shapefiles provided by the Milwaukee Neighborhood Identification Project (MNIP).¹⁵ These neighborhoods are illustrated in Appendix Figure A6. In the 1990s the project sought to unofficially partition the city into neighborhoods using six criteria: subdivisions, major streets and other physical/natural barriers, community group participation, common housing attributes, historic areas, and resident opinions. These criteria capture many types of local amenities that contribute to neighborhood quality, which in turn attract both homeowners and potential commercial storefronts. The resulting map constructed a total of 190 neighborhoods, roughly corresponding in size and shape to a census tract within the city. The high number of neighborhoods relative to the size of the city, as well as the criteria used in their determination, makes the MNIP definitions attractive as a spatial control. Similarly, within the literature, school districts are one of the most commonly researched public goods in relation to housing choice. As such, high school district fixed effects provide an additional spatial control to minimize omitted variables.

Socioeconomic Data

To test the hypothesis that grocery stores have a different marginal implicit price in neighborhoods with a high proportion of minorities, we require demographic data on the ethnic composition of the city of Milwaukee. We use shapefiles and public U.S. Census data for this. We use data at the tract level in this study, specifically racial composition based on its very close mapping to the unofficial neighborhoods in the city.

Estimating the Average Effect of Grocery Stores on Home Prices

We first estimate the average effect of a grocery store on home prices. Table 2 reports the results of regressions excluding commercial controls, and using four different sets of spatial controls. The estimates show that grocery stores are positively capitalized in home prices. In particular, for each additional grocery store within 0.5 mi, the mean price increases by 1.1% to 1.25%. For the specifications that include only spatial fixed effects for high school districts, the impact of a grocery store between 0.5 and 1.0 mi is negative. However, with the inclusion of neighborhood fixed effects, these coefficients become positive and significant at the 1% level, with a grocery store increasing home prices by 0.74% in the model including both neighborhood fixed effects and high school district–specific time trends.

Beyond the change in sign for this coefficient, there are other indicators that the addition of neighborhood fixed effects helps control for unobserved neighborhood amenities, particularly in the model that also contains high school attendance zone time trends. In this model, the variation in home prices arises from changes within neighborhoods by high school district and year. Given the fine level of resolution provided by this combination of fixed effects, the possibility exists that the remaining variation would not be enough to provide significant estimates of the coefficients of interest. However, the regression using these controls exhibits improved fit and retains the signs and significance level of the regression containing only neighborhood fixed effects and time dummies. Given this, in what follows we use the model including neighborhood fixed effects and time dummies differentiated by high school district as our preferred specification.

As noted, one cause for concern is the potential for the coefficients on grocery stores to be biased due to the omission of other commercial properties. Grocery stores rarely appear in isolation and, instead, are usually located in clusters of several commercial and retail properties. To account for this potential source of bias, we introduce several controls for commercial land use to our preferred specification: distance to the nearest commercial lot and its square, the percentage of land within 1.0 mi of each home dedicated to commercial use, and the number of bars in halfmile buffers up to a 1.0 mi ring. Given that the direction of potential bias is unclear, the possibility exists that these additions could change the signs of these estimated coefficients. Table 3 presents these regressions.

In comparison to the initial results, the signs of estimates remain the same across all three specifications. However, in comparison to the model without commercial controls, the estimated impact on home prices increases for grocery stores in each buffer ring as more controls are added, suggesting potential downward bias from not including them previously. As column three illustrates, the premium for an additional grocery store within a half mile increases to 1.34%, and that for a grocery store within a mile rises to 0.85%. This supports the hypothesis that grocery stores provide specific amenities that are distinguishable from general access to retail options.

The estimates for the commercial controls also present an interesting narrative, though these results must be seen through an appropriate lens, as they are constructed primarily as a control, and not as actual amenities. In line with the vast majority of existing hedonic studies, the estimates for proximity to nongrocery commercial land uses exhibit a U-shaped effect on home prices. On average, distance to commercial land uses flips from a disamenity to an amenity at 0.32 mi. The estimated coefficients on the proportion of land dedicated to commercial land use within 1.0 mi report as negative models 2 and 3, though they are insignificant at a 10% level. This is likely the result of low variance in the variable at the spatial scale of the preferred specification. Results from model 3 also indicate that bars between 0.5 and 1.0 mi from a home have a negative average effect on home prices, suggesting the potential existence of noise- and crime-related disamenities. It is also possible that the inclusion of bars as a control is better at capturing the effect of commercial density than the commercial percentage variable, due to the coarseness of the construction of the percentage variable.

Estimating Heterogeneous Implicit Prices for Grocery Stores

The results shown in Table 3 suggest a robust positive average marginal effect for grocery stores within close proximity of a home. To estimate how grocery stores capitalize into home prices differentially across neighborhoods, we interact the number of grocery stores within a given distance ring with the combined percentage of AA and LA households within the homes’ census tract. Table 4 reports the results of these regressions, where we progressively add commercial property controls to establish robustness.

The first column (model 1) presents a regression without any of the commercial land use controls. The estimated average effects for a grocery store remain positive, but they are not significant at the 10% level. However, the interaction term for grocery stores within a half mile is positive and significant at a 10% level, suggesting a positive relationship between the premium for access to a grocery store and the racial composition of a given neighborhood in Milwaukee.

The next two columns highlight the results from models 2 and 3, which use additional commercial land use controls. We use the same controls as in the previous section but also include interaction effects for the number of bars within each buffer ring of a home transaction. As in the previous section, the inclusion of commercial controls increases both the level and significance of the coefficients of interest. The last column presents the results of model 3, which includes both general commercial controls and specific controls for nearby bars. The interaction terms show that prices for a grocery store rise 1.71% and 2.25% for a grocery store within 0.5 mi and 0.5 to 1.0 mi, respectively, in a neighborhood with only AA and LA households.

Table 5 shows the predicted premiums for an additional grocery store within each buffer ring, based on the percentage of AA LA households in their neighborhood, both as a percentage and as a dollar amount based on mean home prices in Milwaukee. Dollar premiums rise to $2,460 for a grocery store within 0.5 mi, and $1,870 for a grocery store within 0.5 to 1.0 mi consisting solely of AA and LA families. Figure 2 illustrates how the dollar premium for a grocery store within 0.5 mi of a home varies across neighborhoods in Milwaukee during a representative year.¹⁶ These premiums rise to as high as $3,215 in some predominately AA neighborhoods. This figure provides a particularly compelling illustration of racially differentiated food access within the city. In contrast with the percentage premiums estimated via regression, these dollar estimates rely on the price paid for homes within each neighborhood, and allow for a more straightforward comparison of the marginal prices paid for an additional grocery store across the city. While the estimates establish that there are differences in the percent premium paid by households, it is possible that the dollar premiums could be equal across the city, or even higher in areas with a lower percentage of AA and LA households. While this would not suggest that food inequality does not exist within the city, this figure illustrates both the higher percentage and dollar premiums paid for an additional grocery store based on the racial composition of a given neighborhood, providing strong evidence of potential food insecurity within the city.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Dollar Premiums for an Additional Grocery Store within 0.75 to 1.0 mi of a Home in 2005

Robustness Check: Including Recession Year Sales

In the analysis provided above we do not include residential transactions from the Great Recession years, as there is evidence to suggest the Milwaukee housing market was not in equilibrium during those years. To test the robustness of our results, and to assess the degree to which these “disequilibrium” years affect the inference of this analysis, we rerun our two preferred specifications of the hedonic model. The first regression is the average effects model including commercial controls, and the second contains the interaction effects model, also with commercial controls. The results are presented in Appendix Table A1. Qualitatively, the results are identical to the corresponding models excluding the recession years, and nearly identical quantitatively. On average, grocery stores are positively capitalized into home prices, increasing home prices from 0.57% to 0.87%. Once interaction terms are added, the level effects for grocery stores lose their significance, but the interaction terms indicate a positive correlation between premiums for grocery stores and neighborhood racial composition. In neighborhoods with only households of color, an additional grocery store within 0.5 mi of a home increases its price by 2.03%, and one within 0.5 to 1.0 mi increases its price by 1.90%. Despite the near identical results, we elect to use our model excluding the recession year results as the primary model, due to the likelihood that the equilibrium assumption was violated during these years.