Abstract
The US retail industry has undergone enormous restructuring resulting in construction of new retail space, abandonment of nearby space, bankruptcies, mergers and acquisitions. This paper estimates discrete choice models of opening and closing probabilities of anchors at a given time and location. A probit model with location fixed effects estimates opening and closing probabilities over time and a conditional logit model (CLM) estimates the odds that a given location will be chosen over a competitor. Probabilities are evaluated from the perspective of a given type of anchor classified as low-, mid- or high-price. New findings include the trade-off between competition from same type anchors and localization benefits (different type) associated with comparison shopping in a retail cluster. We develop a new tool for evaluating risks to any existing retail cluster: risks associated with opening a new anchor and with closures of existing anchors. We demonstrate out-of-sample predictive accuracy.
This is a preview of subscription content, log in via an institution to check access.
Similar content being viewed by others
Notes
Recent studies of the impact of big anchors on smaller retailers include Basker (2005), Ellickson and Grieco (2013), Foster et al. (2006), Haltiwanger et al. (2010), and Hausman and Leibtag (2007). In addition to retail sector, opening/closing decisions of anchors are important to housing and labor market (Pope and Pope 2015; Neumark et al. 2008.
The 2007 Economic Census reports 1,122,703 retail establishments in the United States and a total of 14.2 billion square feet of retail space.
Deadmalls.com contains a partial list of troubled or abandoned shopping centers. When last accessed on March 20, 2014, there were over 450 such stories logged during the period covered by our data (2005 through 2011).
Our discrete choice and risk analysis models are developed specifically for US suburban shopping districts. They are relevant to other countries in two ways. First our methods can model what might happen if controls on retail development are relaxed in response to the needs of local populations. Second, downtown retail outside the US is often very dynamic. For example, Cheshire et al. (2013) analyze the impact of retail development restrictions in the UK on downtown shopping. Our models can be adapted to downtown retail.
“Automobile row,” the clustering of competing automobile dealerships, provides a well-known example of localization.
Our model mimics the decision making of institutional investment funds, where funds are allocated over time to particular metropolitan areas, followed by a choice of a location within that area. This modeling choice is based primarily on numerous presentations of the investment process by professionals at TIAA-CREF, GE Capital, Cornerstone Realty Advisors and other large institutional players.
Source: http://www.kohlsrealestate.com/newcriteria.htm , last accessed April, 2015.
Source: http://articles.courant.com/2012-08-06/business/hc-kohls-connecticut-jobs-expansion- 20120803_1_kohl-s-corp-berlin-kohl-job-fairs, last accessed April, 2015.
Many high-price closings were due to cannibalization after a wave of bankruptcies and mergers in retail industry: e.g., Macy’s acquired May’s department stores in 2007.
Our reduced form findings do not prove space filling strategies, and we remain agnostic as to a game theory model. Clearly closures responded to the GFC shock. However, we find it difficult to construct an explanation of closures other than a myopic space-filling strategy.
The store closing analog to (1) gives p c is,t , the probability of a closing by store type i at location s at time t. The analogous ratio of probabilities is the odds of a closing at the subject site relative to alternatives.
Our competitive condition variables follow Vitorino (2012) and Clapp et al. (2015). Our demand variable follows Clapp et al. (2014) and Zhou and Clapp (2015). Schuetz et al. (2012) suggest that retail patterns vary by neighborhood income. According to Business Insider, 90 % of Americans live within 15 miles of a Wal-Mart.
Source: http://www.businessinsider.com/crazy-facts-about-walmart-2012-11?op=1
Please refer to CRZ (2015) for detailed explanations.
The probit model is further connected to the cross-sectional logit model in this paper: probit predictions of opening probabilities over time by county are used to weight Census tracts considered by the logit model.
Zhou and Clapp (2015) define anchors as department stores where shoppers can find different categories of products including clothing, footwear, bedding, furniture, jewelry, beauty products, and housewares, as well as different brands within each category. Their definition is consistent with the hypothesis of Konishi and Sandfort (2003) that consumers with preference uncertainty can economize their travel costs. Following the same logic, discount stores, hypermarkets and wholesale clubs play a similar function and are normally anchor stores for large power centers and lifestyle centers. Zhou and Clapp (2015) require a minimum size of 20,000 sq.ft. for a typical store. We end up with 49 department store chains in 23 MSAs, 14 MSAs in East and 9 in Central.
See CRZ (2015) for detailed discussions. Clapp et al. (2014) also discuss the reason for excluding downtown shopping.
We do not include announcements of country-wide closures due to bankruptcies at corporate level because many of these stores are acquired by other chains. But we include closings by distressed companies prior to their bankruptcies. A typical example is Mervyn’s, which started to reduce its store in 2007 and announced that it would liquidate its assets and close all the stores by the end of 2008. In this case, we count closures prior to 2008 but not stores acquired by Macy’s and other chains as part of the liquidation. These rebranding activities remain in our database as being operated by the new chain.
The scale of retail change is indicated by openings as a percentage of existings, the gross increase in numbers of anchors. In an average year this was about 3.6 % in the East and 5.3 % in the Central region. For closings, in an average year this was about 0.87 % in the East and 1 % in the Central.
Our use of pre-existing anchors to indicate the feasibility of adding another follows Duranton and Overman (2005).
The effects of the GFC on Wal-Mart and Target are of particular interest because together they average over 40 % of openings and industry professionals credit them with introducing new distribution, logistical and supply management systems that substantially reduced the cost of retailing. The number of openings reduced from 35 (in 2005–2007) to 10 (in 2009–2011) for Wal-Mart and 29 to 10 for Target.
The use of three-mile radius around each tracts centroid is justified by discussions with real estate professionals and by the fact that larger areas would potentially comprise too much of a typical metropolitan area, depriving location choices of sufficient intra-metropolitan variation. Fanning (2005, page 192) uses three to five miles or a 5–10 min drive on local roads as the boundary between shopping centers anchored by supermarkets and junior department stores and those anchored by the multiline department stores we study. Moreover, a three-mile radius is consistent with the sizes of submarkets defined by CoStar and other vendors. Moreover, three-mile areas have the advantage of encompassing enough area to obviate the need for distance weighted variables such as those used by Crozet et al. (2004) and by Woodward et al. (2006). Spatial autoregressive techniques are not well developed in the CLM context. Despite the advantages of using three-mile LAs, we are aware of the Modifiable Areal Unit Problem (MAUP) and so we provide sensitivity tests of the three mile assumption. We repeat our tests by using 1-mile, 2-mile and 5-mile and conclude similar results.
Each opening is paired with five rejected alternatives (Rej_O), LAs without openings by the chain within the same MSA. Similarly, “Closing” signifies the LA of a closing. Given a closing, the rejected alternatives (Rej_C) are LAs with pre-existed stores of the same company but without closings. It means we have more candidates for rejected alternatives of openings than that of closings. In order to maintain a comparable number of counterfactuals, we do not require them to be within the same MSA as the closing, an assumption reflecting the reality of decisions about which stores to close. Details of LA classifications and variable definitions are included in Table descriptions.o
As there is no accurate time-vary demand side measurement at tract level, we first estimate potential demand (number of household times median household income then divided by total land areas) as of the beginning of our sample period, 2005 and then multiply it by the cumulative percentage change of aggregate payroll in the county where the LA is located from 2005 to year t. We assume that all the LAs within the county experience homogenous shocks. As there is no time-vary demand measures at tract level, there is a trade-off between time-varying and cross-sectional accuracy. Zhou and Clapp (2015) estimate expected demand as of 2010 by multiplying demand in 2000 by average annual growth rate from 1990–2000. As a result, they apply a static measure to all the LAs and there is no GFC effect on their demand variables. Our method is more advanced compared with Zhou and Clapp (2015) because we allow demand varying over time as well as over space. In addition, we take the GFC effect into consideration.
A full description of CLM model is in Zhou and Clapp (2015). An important addition here is that locations considered for openings (“counterfactual” locations0 to be weighted by probabilities from the probit model, so that a county with high probability over time is more likely to be chosen when estimating the CLM.
We include highway access as Holmes (2011) suggests that Wal-Mart’s expansion relies heavily on distribution centers, which are built along highways (see “the Wal-Mart model of a big box store at a convenient highway exit” on page 260). In addition to sales revenue, highway access is an important variable of profit maximization due to logistic benefit and cost reduction.
In unreported results, our findings are robust (1) when we apply a similar method as in probit and keep only counties with at least one existing store as of 2005, instead of all the counties, in our choice set of openings; (2) when we keep only counties with at least one existing store as of the beginning of the year of openings (as compared to the beginning of 2005) in our choice set, and (3) when we use a 5-mile radius to define LAs.
There is no high-price opening in 2011, as shown in Panel A of Table 1.
CLM allows different weights across group but not within group. As a result, we impose different weights when we select rejected alternatives.
The choice set of openings is different from that of closings. For example, in order to be included in the choice set of closing, the LA has to have at least one existing store prior to closing. The choice set also varies across different price types. For example, the rejected alternatives of openings are selected within a MSA. If there is no high-price opening within a given MSA, LAs in that MSA would be selected as rejected alternatives of high-price openings. We argue that this method, as opposed to selecting all the LAs for each model, would run against us finding the statistical significance in Table 5.
As stressed in “The Model”, the probabilities used to calculate the odds ratios could come from expert opinion instead of the model. However, this section will demonstrate the relevance of the variables considered in the model.
Specialty retailers offering choice within this retail cluster include Lowe’s Home Improvement, Super Stop&Shop, Bobs Discount Furniture and Barnes and Noble. Lincoln Plaza was renovated in 2004.
If, counterfactually, we move A2 to within 3 miles of the CBD, its probability drops by 75 % and the odds of locating a low-price anchor at Lincoln instead of A2 increase to over 4. The judgement of a local retail expert might modify the sharp 3 mile boundary effect present in our model.
As with openings, we re-estimate these models over 2005–2010 and predict to 2011.
We cannot compare predicted probabilities among the three chains based on all the existing stores because they have different numbers of existing stores (i.e. the sum of exponential terms of Xbetas, the denominators, are different). But we can compare predicted probabilities among chains with the same number of randomly selected alternatives. Based on the same number of alternatives, predicted probabilities of Kohl’s closing are comparable to those of TJ Maxx because Kohl’s have less same-price existing stores, more high-price stores, are less likely to locate within 3 miles from the CBD but have larger distance to CBD.
The median (mean) of predicted closing probability of the 164 Kohl’s stores is 0.005 (0.006), compared with 0.003 at the SS. The median (mean) of predicted closing probability of the 153 TJ Maxx stores is 0.005 (0.006), compared with 0.0012 at the SS. The median of predicted closing probability of the 189 Target stores is 0.002 (0.005), compared with 0.00018 at the SS.
LAs of low-price (mid-price) stores have the median numbers of mid-price stores of one (two). As a result, the dummy variable of mid-existing equals zero (one) in closing-low (closing-mid) model in Table 5 Model (4) (Model(5)).
References
Basker, E. (2005). Selling a cheaper mousetrap: Wal-Mart's effect on retail prices. Journal of Urban Economics, 58, 203–229.
Cheshire, P., Hilber, C., Sanchis-Guarner, R. (2013). towards estimating the carbon footprint of retail: do english planning policies make shops more accessible? Working Paper.
Chung, C. J., Inaba, J., & Leong, S. T. (2002). The Harvard design school guide to shopping / Harvard design school project on the City. Harvard University Graduate School of Design. Taschen.Crozet, M., Mayer, T., Mucchielli, J.-L., 2004. How do Firms Agglomerate? A Study of FDI in France. Regional Science and Urban Economics, 34, 27–54.
Clapp, J., Salavei, K., & Zhou, T. (2014). Expansions and contractions of major US shopping centers. Journal of Real Estate Finance and Economics, 48, 16–56.
Clapp, J., Ross, S., Zhou, T. (2015). Retail Agglomeration and Competition Externalities: Evidence from US Openings and Closings of Multiline Department Stores. Working Paper.
Crozet, M., Mayer, T., & Mucchielli, J. (2004). How do firms agglomerate? A study of FDI in France. Regional Science and Urban Economics, 34, 27–54.
Duranton, G., & Overman, H. (2005). Testing for Localisation using Microgeographic data. Review of Economic Studies, 72, 1077–1106.
Ellickson, P. B., & Grieco, P. L. E. (2013). Wal-Mart and the geography of grocery retailing. Journal of Urban Economics, 75, 1–14.
Fanning, S., (2005). Market analysis for real estate: concepts and application in valuation and highest and best use. Appraisal Institute.
Fernandez-Val, I. (2009). Fixed effects estimation of structural parameters and marginal effects in panel probit models. Journal of Econometrics, 150, 71–85.
Foster, L., Haltiwanger, J., & Krizan, C. (2006). Market selectin, reallocation, and restructuring in the U.S. retail trade sector in the 1990s. Review of Economics and Statistics, 88, 748–758.
Haltiwanger, J., Jarmin, R., & Krizan, C. J. (2010). Mom-and-Pop meet big-box: complements or substitutes? Journal of Urban Economics, 67, 116–134.
Hausman, J., & Leibtag, E. (2007). Consumer benefits from increased competition in shopping outlets: measuring the effect of Wal-Mart. Journal of Applied Econometrics, 22, 1157–1177.
Holmes, T. (2011). The diffusion of Wal-Mart and economies of density. Econometrica, 79, 253–302.
Jia, P. (2008). What happens when Wal-Mart comes to town: an empirical analysis of the discount retailing industry. Econometrica, 76, 1263–1316.
Konishi, H., & Sandfort, M. (2003). Anchor stores. Journal of Urban Economics, 53, 413–435.
Kramer, A., (2008). Retail development handbook (Development Handbook Series), fourth ed. Urban Land Institute.
Neumark, D., Zhang, J., & Ciccarella, S. (2008). The effects of Wal-Mart on local labor markets. Journal of Urban Economics, 63, 405–430.
Pope, D. G., & Pope, J. C. (2015). When Walmart comes to town: always low housing prices? Always? Journal of Urban Economics, 87, 1–13.
Salvo, A. (2010). Inferring market power under the threat of entry: the case of the Brazilian cement industry. Rand Journal of Economics, 41, 326–350.
Schuetz, J., Kolko, J., & Meltzer, R. (2012). Are poor neighborhoods “retail deserts”? Regional Science and Urban Economics, 42, 269–285.
Vitorino, M. (2012). Empirical entry games with complementarities: an application to the shopping center industry. Journal of Marketing Research, 49, 175–191.
Woodward, D., Figueiredo, O., & Guimarães, P. (2006). Beyond the Silicon Valley: University R&D and High-technology location. Journal of Urban Economics, 60, 15–32.
Zhou, T., & Clapp, J. (2015). The location of new anchor stores within metropolitan areas. Regional Science and Urban Economics, 50, 87–107.
Acknowledgments
We would like to thank David M. Geltner (the editor), an anonymous referee, S.E. Ong, Joseph Ooi, Rogier Holtermans, Piet Eichholtz, and participants at the Maastricht–NUS–MIT Symposium 2014. We are indebted for excellent GIS assistance from Joseph Danko.
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendices
Appendix 1
Openings and closings by MSA
MSA | Region | Number of openings | Number of closings |
|---|---|---|---|
Albany-Schenectady-Troy, NY | East | 7 | 4 |
Allentown-Bethlehem-Easton, PA-NJ | East | 16 | 1 |
Birmingham-Hoover, AL | Central | 23 | 5 |
Bridgeport-Stamford-Norwalk, CT | East | 4 | 3 |
Cincinnati-Middletown, OH-KY-IN | Central | 29 | 7 |
Columbus, OH | Central | 24 | 7 |
Dayton, OH | Central | 8 | 7 |
Detroit-Warren-Livonia, MI | Central | 64 | 20 |
Grand Rapids-Wyoming, MI | Central | 11 | 1 |
Hartford-West Hartford-East Hartford, CT | East | 6 | 1 |
Indianapolis-Carmel, IN | Central | 21 | 9 |
Jacksonville, FL | East | 35 | 4 |
Louisville-Jefferson County, KY-IN | Central | 15 | 2 |
Milwaukee-Waukesha-West Allis, WI | Central | 24 | 3 |
Nashville-Davidson--Murfreesboro, TN | East | 28 | 7 |
New Haven-Milford, CT | East | 11 | 1 |
Orlando-Kissimmee, FL | East | 48 | 4 |
Providence-New Bedford-Fall River, RI-MA | East | 15 | 3 |
Raleigh-Cary, NC | East | 28 | 1 |
Richmond, VA | East | 19 | 5 |
Rochester, NY | East | 15 | 2 |
Virginia Beach-Norfolk-Newport News, VA-NC | East | 27 | 4 |
Worcester, MA | East | 10 | 4 |
Appendix 2: Variable Definitions and Technical Details of the CLM
Panel A Time-Series Analyses—Fixed Effect Probit with Bias Correction
Variable | Definition |
Dependent variable(s) | |
Open_Low-price | A dummy variable equals one if there is any low-price anchors opened within the county, and zero otherwise |
Open_Mid-price | A dummy variable equals one if there is any mid-price anchors opened within the county, and zero otherwise |
Open_High-price | A dummy variable equals one if there is any high-price anchors opened within the county, and zero otherwise |
Close_Low-price | A dummy variable equals one if there is any low-price anchors closed within the county, and zero otherwise |
Close_Mid-price | A dummy variable equals one if there is any mid-price anchors closed within the county, and zero otherwise |
Close_High-price | A dummy variable equals one if there is any high-price anchors closed within the county, and zero otherwise |
Independent Variables | |
# Low-existing | Number of low-price anchors pre-existed within the county at the beginning of the year |
# Mid-existing | Number of mid-price anchors pre-existed within the county at the beginning of the year |
# High-existing | Number of high-price anchors pre-existed within the county at the beginning of the year |
Demand-AP | Lagged log of aggregate payroll in all sectors. Data on aggregate payroll is collected from County Business Pattern. |
Panel B Cross-sectional Analyses—Conditional Logit Model
Variable | Definition |
Dependent Variable(s) | |
Open | A dummy variable equals one if there is any anchors opened given the local market area (LA), and zero otherwise |
Close | A dummy variable equals one if there is any anchors closed given the local market area (LA), and zero otherwise |
Independent Variables | |
# Low-existing | For openings, it is the number of low-price anchors pre-existed within the LA at the beginning of the year of opening. |
For closings, it is a dummy variable that equals one if the number of low-price existing stores in the LA is greater than the median number of low-price existing stores for the company at the beginning of the year of closing, and zero otherwise. | |
# Mid-existing | For openings, it is the number of mid-price anchors pre-existed within the LA at the beginning of the year of opening. |
For closings, it is a dummy variable that equals one if the number of mid-price existing stores in the LA is greater than the median number of mid-price existing stores for the company at the beginning of the year of closing, and zero otherwise. | |
# High-existing | For openings, it is the number of high-price anchors pre-existed within the LA at the beginning of the year of opening. |
For closings, it is a dummy variable that equals one if the number of high-price existing stores in the LA is greater than the median number of high-price existing stores for the company at the beginning of the year of closing, and zero otherwise. | |
Demand | Log of potential revenue in the LA in the year prior to the opening/closing. |
Potential revenue at tract-level in 2005 is estimated based on per-square-mile number of household (2000 Census) multiplied by median household income (2000 Census) as of 2000 and annual growth rate from 1990 to 2000. Time- varying potential revenue at tract-level after 2005 is calculated based on potential revenue at tract-level in 2005 (in constant 2005 dollar) and cumulative percentage change of aggregate payroll in the county where the LA is located from 2005 to the year of the opening/closing. Data on aggregate payroll is collected from County Business Pattern. | |
CBD_3mile | A dummy variable equals one if the LA is within three miles of the CBD, and zero otherwise. |
hwy_half_mile | A dummy variable equals one if the LA is within half miles from highway, and zero otherwise. |
hwy_half2two_mile | A dummy variable equals one if the LA is from half miles to two miles from highway, and zero otherwise. |
dis2CBD | Distance from the centroids of the LA to the CBD of the MSA |
distance to HQ | A dummy variable equals one if, for a given chain, the distance from the centroid of the LA to its headquarter is greater than the median distance to the headquarter from all the pre-existed stores within the same MSA as of the beginning of the year of opening/closing, and zero otherwise. Headquarter relocations are taken into account. |
distance to DC | A dummy variable equals one if, for a given chain, the distance from the centroid of the LA to its nearest distribution center is greater than the median distance to its nearest distribution center from all the pre-existed stores within the same MSA as of the beginning of the year of opening/closing, and zero otherwise. |
Openings and closings of distribution centers, as well as their usage (for example E-Commerce versus retail sale), are taken into account. | |
Panel C Explanation of Xbeta and Confidence Range in the Conditional Logit Model
Utility maximization in the CLM model implies that the probability of an anchor of a given type locating at the subject site can be calculated from the CLM coefficients:
where p o is is the probability that anchor type i will open at in local market s; i = 1, … N represents opening decision makers (e.g., a low-price anchor) and j = 1, … J represents a range of alternative locations. The set of J alternatives includes the subject site, s, and x j is a vector of location-specific variables which does not depend on decision makers. β x is a vector of unknown parameters of x j . A vector of location-specific variables which depends on decision makers is given by z ij . β z is a vector of unknown parameters of z ij . Equation (A1) can be written as \( {p}_{{}_{is,t}}^o=f(Xbeta) \).
The parameters are a linear approximation to a reality that is more complex. Any set of parameters is more informative if the planner can obtain a confidence range. We accomplish this by shrinking an individual parameter estimates towards zero by one standard deviation and expanding if away from zero by one standard deviation:
where \( \widehat{\beta} \) without a prime or a subscript indicates an individual element of the vectors β ′ x or β ′ z . We choose one standard deviation because the normal approximation implies that this will capture about two-thirds of the true parameter values. This leads naturally to odds ratios based on low and high parameter estimates, Odds(s, j; i)l and Odds(s, j; i)h. Note however, that it is not necessarily the case that Odds(s, j; i)l < Odds(s, j; i)h. The local planner can judge relative probabilities of opening for any anchor-type i at alternative sites j using this range of estimated odds ratios for each alternative j that is competitive with the subject site (i.e. comparing Odds(s, j; i)l (odds of low Xbeta), Odds(s, j; i)h (odds of high Xbeta) and odds for most likely Xbeta.
Appendix 3: All Scenarios of Time-Series Analysis of Openings in Worcester Application
Note: This figure shows all scenarios of time-series analysis of openings in Worcester Application. The numbers under the first node, “t”, represents probabilities of low, mid and high-priced openings at t, where “L: (0.074, 0.083)” means the probabilities of low-priced openings at the SS when parameters imply low odds (i.e. 0.074) and high odds (i.e. 0.083). The second set of nodes, “Low”, “Mid” and “High”, represents probabilities of low, mid and high-priced openings in t + 1 under different scenarios. For example, the “Low” node means there is a low-priced anchor opening at the beginning of t + 1. “L: (0.048, 0.054)” shows the probabilities of the low-priced openings at the SS imply low odds and high odds, given a low-priced opening at the beginning of t + 1. The third set of nodes, three “Low”, three “Mid” and three “High”, represents probabilities of low, mid and high-priced openings in t + 2 under different scenarios. For example, the first “Low” node means there is a low-priced anchor opening at the beginning of t + 2, given a low-priced anchor opening at the beginning of t + 1. “L: (0.031, 0.035)” shows the probabilities of the low-priced openings at the SS under low odds and high odds, given a low-priced opening at the beginning of t + 1 and another low-priced opening at the beginning of t + 2.
Rights and permissions
About this article
Cite this article
Zhou, T., Clapp, J.M. Predicting Risks of Anchor Store Openings and Closings. J Real Estate Finan Econ 52, 449–479 (2016). https://doi.org/10.1007/s11146-015-9524-1
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11146-015-9524-1
Keywords
- Localization economies vs competition
- Multi-line department stores
- Retail locations
- Conditional logit
- Anchor stores
- Retail opening and closing