Research Questions and Empirical Strategy

The past half century of economic research on the capitalization of residential housing prices of different neighborhood amenities and disamenities associated with nearby land uses and access to features of the landscape demonstrates that homebuyers have preferences over these features. Most of these analyses use a hedonic pricing approach to test hypotheses concerning the sign and size of these effects. Hedonic models are reduced-form statistical models used to trace the locus of equilibrium transaction prices as a function of the characteristics of the real estate transacted. These models assume that a residential property is a heterogeneous good, made up of a bundle of characteristics. Each characteristic of the parcel, including surrounding land uses and access to landscape features, contributes to the sales price of the property. Examples of applications of hedonic models to value different amenities and disamentities include: clean air (Smith and Huang 1993); clean water (Leggett and Bockstael 2000); open space (Geoghegan 2002); Superfund sites (Kiel and Williams 2007); airport noise (Pope 2008); transportation access (Bowes and Ihlanfeldt 2001); among many others (for an overview of hedonic property models, see Palmquist 2005).

Early examples that explicitly try to measure the effect of commercial land uses on neighboring housing prices include Li and Brown (1980) and Song and Knapp (2004). In two recent applications (Johnson et al. 2015, Pope and Pope 2015), researchers investigate the effect of large “big box” retail establishments, such as Target and Walmart on nearby residential land markets. While Walmart is currently the largest grocery retailer in United States and as a result can have an impact on local food prices, (Ellickson and Grieco 2013), these stores also are much larger than a “traditional” grocery store, sell many other household items, and do not often open in urban core neighborhoods. Therefore the results from these studies may not carry over to our research question. However, the key motivations and empirical strategy put forth by Pope and Pope (2015) are relevant for us. They argue that a new Walmart could be viewed an amenity, valued for increased ease of access, or as a disamenity, due to congestion and other negative externalities. We make a similar argument for the impact of a new grocery store on nearby residential values as well, specifically that the grocery store is a disamenity for properties that are very close, as a result of congestion and noise externalities, while the grocery store becomes an amenity at a slightly further distance, because of easier access, but that the effect decays as the distance between the house and the grocery store increases. However, it is an empirical question of the relevant distances to have a measureable impact, and therefore we use a number of different distances to test this effect in our empirical specification.

Pope and Pope (2015) also discuss the limitations of the early hedonic models that investigated issues related to the impact on housing prices of different externalities and accessibility features of the landscape. First, these models used a cross-sectional approach to estimate the effect of the feature in question; therefore, omitted variable bias could affect the estimated implicit prices if other features of the landscape also influence residential land prices but are not included in the model. Second, by ignoring the endogenous nature of the locational choice of a new store, a traditional cross-sectional approach could also result in biased estimation of implicit prices. Instead, Pope and Pope advocate for the use of new methods in quasiexperimental design that make use of discontinuities in time and/or space to help clearly identify and estimate the implicit prices of the characteristics of interest (for an overview of this approach, see Parmeter and Pope 2013; for an application of this approach to housing markets in Worcester, see Brown and Geoghegan 2011). Given information on both the timing and location of new Walmart stores opening, they develop a difference-in-differences approach to compare residential sale prices before and after the construction of a Walmart store for residential properties located near a new store (the treatment group) to residential properties located further away from a new store (the control group). The key feature of this empirical strategy is that it can avoid the potential endogeneity of both the location and timing decision of the opening of these new stores.

Given similar concerns for the endogeneity of the timing and location of new grocery store openings in Worcester, we take a parallel difference-in-differences approach as well, using temporal and spatial discontinuities to address the potential endogeneity issue. As Pope and Pope argue, by including spatial fixed effects in the empirical specification and comparing residential sales before and after the opening of each new grocery store, the omitted variables that could potentially bias a traditional hedonic approach that do not change over time are differenced away. However, for this approach to be valid, the assumption is that the time trends in housing prices would be the same in the treatment and control groups would have been the same without the opening of the new store (Pope and Pope, 2015). Given that our analysis focuses on a single city, coupled with the spatial specificity and disaggregation of our data, we believe this is a reasonable assumption. In addition, we include falsification tests and time trend analysis below to further justify this assumption.

Therefore, similar to Pope and Pope, we use residential properties within an 800-meter (approximately one-half mile) radius of the grocery store as the treatment group, and properties located more than 800 meters away as the control, with interactions with an indicator variable for properties that sold after the new store opening ¹ . Our empirical specification is:

(1) $${\rm ln}\lpar{Price_{ijy}} \rpar= \; {\rm \beta}_{\,jy} + \; {\rm \alpha} {\bi X}_i + {\rm \gamma} {\bi Y}_i + {\rm \theta}_1^d {\bi D}_{ij}^d + \; {\rm \theta}_2^d {\bi D}_{ij}^d {\bi Post}_{iy} + {\rm \varepsilon}_{ijy} $$

The log of a sales price of home i is a function of a grocery store j by year y; a specific store and year effect (β _jy ); an indicator variable for the year of residential sale (Y _i ); observable individual house characteristics as well as neighborhood characteristics (X _i ); indicator variables, ( $D_{ij}^d$ ) for sales within a prespecified distance (d) from the grocery store regardless of whether the grocery store has been established or not (where the distance varies as described below in the data description); interactions of each of these spatial indicators with an indicator (Post _iy ) for if the sale occurred after the opening of the grocery store, and an error term (ɛ _ijy ).

The estimated parameters of the model include: α is a vector of coefficients for the effects of various housing and neighborhood characteristics; γ captures annual changes in the real estate market that could result from macroeconomic fluctuations; θ₁ estimates the fixed effect of being located within distance d from a grocery store; and θ₂ measures the price effect of being located within distance d from a grocery store after the date of opening of the grocery store. Statistical significance of θ₂ implies that grocery store openings have an effect on a particular neighborhood's real estate market. In addition, as discussed above, we hypothesize that the coefficient will be negative for properties located adjacent to, or in close proximity to, a grocery store, because of congestion and other negative externalities, but positive for properties located somewhat further away, due to the positive effect of accessibility, but that the positive impact is expected to fall as the distance between the new grocery store and the residential property increases.

Given that the timeframe of our study, from 1988 to 2011, includes a dramatic increase in housing prices that occurred throughout Massachusetts during the middle of the timespan (Goodman et al. 2004), as well as the crash in the housing market associated with the Great Recession of 2008, both the level and time trend in prices are critical considerations in our application. While it does not appear that there is a great deal of difference between the overall time trend in housing prices for Worcester as a whole compared to the subgroup of housing sales within approximately one-half mile of all new grocery stores as can be seen in Figure 1, with further spatial disaggregation, the time trends can vary more greatly. Using a specific grocery store as an example, Figure 2 shows the time trend for all residential sales over the time period within 800 meters of this particular grocery store compared with residential sales within more disaggregated distances around the grocery store. There is more variability in housing prices at this particular scale in both the temporal and spatial dimensions. Therefore, similar to Pope and Pope (2015), we check the validity of our difference-in-differences results by exploring a set of falsification tests in which the opening year of grocery stores is shifted to earlier years.

a) Includes all properties in the sample b) Includes properties located within half a mile from a grocery store.

Figure 1. Median Price of Houses.

a) Includes all properties 0–800 meters b) Separates properties by buffering zone

Figure 2. Average Price of Houses Located 0–800 Meters from Price Chopper Located at 50 Cambridge Street, Worcester MA.

Data Description

Housing sales data from the Warren Group were purchased for all residential sales that occurred in the city of Worcester over the period 1988 through 2011. These data include information on the location of the property, date of sale, style of the structure, lot size, and other housing characteristics. Neighborhood characteristics are from the American Community Survey (ACS) 5-year summary (2006–2010). Data on grocery store location and year of opening were obtained from Reference USA (July 2012) for SIC code equal to 541105 and 531102. The first SIC code corresponds to grocers-retail businesses, and the latter is for supercenters that also sell food products, of which there was one in our sample, at the southern edge of the city (see Figure 3). Date of grocery store opening was confirmed through Worcester City's directories for assorted years (Polk 1988–2011). During our study period, twelve new grocery stores opened. The stores were at least 40,000 square feet in size and sold a wide variety of food products ranging from fresh vegetables and fruits to processed foods. Excluded from the data were smaller specialty stores, convenience stores, and retail bakeries. Figure 3 shows the evolution of grocery stores in Worcester during the study period.

Note: Grocery stores opening on 1987 or before are labeled as “1987.”

Figure 3. Years of Grocery Stores Opening in Worcester.

All these data were geocoded for spatial analysis. Two mutually exclusive spatial buffers were created around each of the twelve new grocery stores (0–399.9 meters and 400–800 meters), representing the “d” distances for the indicator variables discussed in equation 1 to determine the residential sales within each distance category. The summary statistics for the housing data, including sales price, and neighborhood characteristics can be found in Table 1.

Table 1. Summary Statistics of Housing Data

Results and Discussion

Our analysis follows the difference-in-differences approach presented in Equation 1. The regression includes spatial indicators to distinguish prices between residential sales properties located more than 800 meters from each new grocery store to properties located within 800 meters. The properties within the 800-meter distance were divided into two mutually exclusive rings which allow for a more detailed comparison between homes located close to a grocery store to those located further away. The estimated coefficients and standard errors of these spatial variables and their interactions with the post-opening grocery store indicator are shown in Table 2. Column 1 includes all the properties in the sample, column 2 excludes properties that had a sales price in the top 1 percent of sales price observations, and column 3 excludes properties that had a price in the top 5 percent of sales price observations. Standard errors were clustered at the blockgroup level. The estimates of the coefficients of the spatial variables suggest that properties located within each of the two spatial buffers were associated with lower prices compared to properties located more than 800 meters away. Each of the indicator variables $D_{ij}^{0 - 399.9} $ and $D_{ij}^{400 - 800} $ has a negative and statistically significant coefficient indicating that properties located within 800 meters from a grocery store site sold at a price between 7 and 8 percent below the value of comparable properties located more than 800 meters away from the store. Overall, this suggests that grocery stores are located in lower priced neighborhoods. The first row of Table 1 shows the average sales price of all the properties sold between 1988 and 2011. Average sales price for properties within the 800-meter buffers were below the city's average sales price and below the prices of properties located more than 800 meters away from a grocery store. This could reflect the differences in housing structure and neighborhood characteristics. Houses within the 800-meter buffers are on average older and have a smaller lot size, and these neighborhoods are more racially diverse and densely populated.

Table 2. Impact of Grocery Store Openings in Housing Prices

Robust standard errors in parentheses

*** p < 0.01, ** p < 0.05, + p < 0.1

Note: All regressions control for year fixed effects

In terms of the impact of a new grocery store development on real estate prices, the estimated coefficient of the interaction term $D_{ij}^d Post\; _{iy} $ identifies the effect on sales price of a house located within a specific distance band from the grocery store. The estimated coefficients of $D_{ij}^{0 - 399.9} Post\; _{iy} $ and $D_{ij}^{400 - 800} Post\; _{iy} $ have positive magnitudes. In addition, the coefficient is larger for properties located closest to the grocery store. For example, as shown in columns 1–3 of Table 2, properties located within 399.9 meters from a grocery store experienced a 7 percent increase in price after the store opened, while properties located within 400–800 meters from the grocery store experienced approximately a 4 percent increase in value. As a comparison, Pope and Pope (2015) found that a new Walmart store increases housing prices between 2 and 3 percent for houses located within 0.5 miles of the store, while similar positive results for a new Walmart within a two-mile radius of a residential prices can be found in Johnson et al. (2015).

The estimated coefficients on the control variables have the expected signs and are statistically significant: homes with more bedrooms and bathrooms sell at a higher price, as do properties with larger lots. Homes in neighborhoods with lower income, lower education levels and higher percentages of minorities sell at lower prices.

In order to test the robustness of these results, we further disaggregated the 800-meter distance into four mutually exclusive rings (0–199.9 meters, 200–399.9 meters, 400–599.9 meters, and 600–800 meters) centered around each of the grocery stores. One of our hypotheses was that a new grocery store would have a nonmonotonic effect on neighboring housing prices, where properties in close proximity to a grocery store experienced a negative impact, and properties located further away experienced a positive distance-decaying effect. However, as shown in Table 3, this relation was not observed across the coefficients of the interaction terms. In fact, the coefficients of the interaction terms $D_{ij}^{0 - 199.9} Post\; _{iy} $ and $D_{ij}^{400 - 599.9} Post\; _{iy} $ are not statistically significant.

Table 3. Impact of Grocery Store Openings in Housing Prices for More Disaggregated Buffers

Robust standard errors in parentheses

*** p < 0.01, ** p < 0.05, + p < 0.1

Note: All regressions control for year fixed effects

Finally, falsification tests were conducted to identify whether the observed increase in housing prices was due to a grocery store opening a short distance from the property. One of the concerns is that there may be confounding factors affecting the price of houses close to the grocery store's site before its announcement and opening date. Hence, similar to Pope and Pope (2015), we constructed false opening dates for the grocery stores and shifted the opening years of grocery stores three, four, and five years prior to the actual opening years ³ . If the grocery store had an impact on property values, we should observe a positive effect when the grocery store is announced and opened, and not in prior years. The results of these tests are shown in Table 4. None of the coefficients of the interaction terms provide evidence of property values near a grocery store's site increasing in value three, four, and five years before the store's actual opening year. Thus, these results suggest that the opening of a new grocery store does have a positive effect on nearby residential property values.

Table 4. Falsification Tests of Grocery Store Opening Year and Impact on Housing Prices

Robust standard errors in parentheses

*** p < 0.01, ** p < 0.05, + p < 0.1

Note: All regressions control for year fixed effects