The Geographical Edge:
Spatial Analysis of Retail Loyalty Program Adoption
Arthur W. Allway, The University of Alabama, USA
Lisa D. Murphy, The University of Alabama, USA
David K. Berkowitz, The University of Alabama, USA
Abstract
This chapter demonstrates important insights gained by adding spatial capabilities to marketing analyses. Four steps are described to produce a geographically enabled data set of the first year’s daily use for a major retailer’s loyalty card program at one store in a mid-western U.S. city. Traditional analysis is contrasted with results from a geographic information system (GIS). Probabilities of adoption were clearly tied to the geographic variables generated by the GIS; for example, over the whole year, the likelihood of someone adopting on a given day decreased 13.4% for each mile they resided away from the store, while each Innovator (adopted in the first two days) located within .6 mile of a prospective adopter increased adoption likelihood by 13.2%. Further, three very distinct spatial diffusion stages are visible showing adoption as a function of distance to the store itself, to the billboards, and to the earliest adopters.
The Geographical Edge 261
Introduction
Today’s retail marketing managers have access to better information than ever before.
In particular, the spread of point-of-sale automation in retail stores has turned what used to be a trickle of data into a flood. For many retailers, this technology has become the basis for the development of innovative, customer-centered loyalty card programs. A battalion of intercept interviewers in a store for weeks or a buyer’s panel operating for months can capture only a small portion of the data gathered by a point-of-sale loyalty card program every day. To make sense of this data deluge, marketers are having to rely on a battery of both familiar statistical techniques such as regression analysis and newer ones such as chaid and diffusion modeling.
Much of the value of the data generated by a POS-based loyalty card program is its ability to capture the speed and duration of market reaction to new store openings, product launches, advertising campaigns, promotions, and so on. As such, loyalty programs often lend themselves to a diffusion of innovations analysis approach. Yet, even though retailing (except web-retailing) is necessarily a geographically anchored activity, diffu- sion research has typically ignored geographic factors. The reasons for this neglect have for the most part been practical. Prior to the advent of geographic information systems (GIS), spatial data was difficult to use, expensive to collect, and often of uncertain quality.
The most common tools for analyzing spatial data were paper maps and overlays — both cumbersome to use and difficult to update and refine. As a result, even marketers who clearly recognized the importance of geography in both their and their customers’
decision-making seldom received the tools or training that would make geography worth addressing at the individual consumer level (Murphy, 1996).
The application of GIS to retail point-of-sale data holds great promise in allowing retailers to gain greater insights into consumer spatial behavior. With that in mind, this chapter attempts to add to the body of retail theory and practice by demonstrating how a GIS- centered spatial approach can expand researcher understanding of the diffusion of a new loyalty card program. Household-level data from the entire first year of a new loyalty program launched by a very large retailer in a major U.S. city is combined with GIS- generated measures to explore the effect of distance, marketing efforts, and other adopters on the diffusion process of consumer adoptions.
This chapter will demonstrate how adding spatial analysis to traditional market innova- tion approaches can help make sense of a huge volume of data, provide insights into the patterns of adoption and the influences on adopters, and ultimately help improve decision-making. Our goal is not to demonstrate the absolute superiority of spatial techniques over other approaches nor to develop new theory about spatial influences on diffusion, but to illuminate for both practitioners and researchers some areas where new insights may await both discovery and application. We consider this to be a particularly relevant goal with the new opportunities presented by having significant actual purchasing data and geographic analysis tools.
The chapter proceeds with an overview of background material followed by a description of this study and its data sources. Keeping with the objectives of this book, the steps required to utilize GIS with this particular data set are described step-by-step. The results
262 Allway, Murphy and Berkowitz
are presented in tables and figures to facilitate the comparison between the insights that would be gained with and without GIS. The chapter concludes with a few implications for researchers and practitioners in this area.
Background
Geographic considerations have been central to the study of retailing. All retail marketing decisions have to take into account their probable impact on the size, shape, depth, and/or dynamics of the market area of the firm. From the consumer side, such personal decisions as willingness to travel, impediments to arrival, relative visibility of location, reaction traffic patterns, and the influence of competitive locations are also geographic in nature.
Three specific research streams have concentrated specifically on the geography of retailing. One research stream has concentrated on delimiting trade area boundaries so that business decisions that affect the sizes and shapes of those market areas can be evaluated more precisely (see, for example, Huff & Batsell, 1977; Donthu & Rust, 1989).
Another stream of geographical research in retailing has involved the modeling of consumer choices in spatially defined markets (see, for example, Huff, 1962, 1964; Ben Akiva & Lerman, 1985). A third, although still emerging, stream is concerned with the spatial diffusion of consumer response to marketing efforts. Although a significant body of spatial diffusion theory does exist in geography and sociology (beginning with Hagerstrand, 1967), little of it has focused on retailing (Allaway, Berkowitz, & D’Souza, 2003).
Diffusion of innovations has proved a useful and durable explanation of how commu- nication affects human behavior. This theory, pioneered by Everett Rogers (1962, 1983, 1995), is based on the notion that a new innovation is first adopted by a few innovators, who, in turn, influence others to adopt it, typically via word of mouth. Continuing influence of adopters on potential adopters explains the shape of the sales trajectory curve over time (Rogers, 1995). Spatial diffusion research adds the geographical element to this research, explaining the patterns of adopter interaction with potential adopters spatially as well as temporally.
Spatial diffusion research in marketing has been hampered in the past by large-scale requirements for spatially coded data, which has been traditionally difficult and time- consuming to acquire and use (Murphy, 1996). However, two new technologies have emerged that make the potential for doing spatial research faster, easier, and more accurate. This paper demonstrates the payoff that bringing these two technologies together can have for both retail practitioners and academics. One of these technologies
— point-of-sale-based customer loyalty programs — delivers vastly improved customer- specific behavior data. The other technology — geographic information systems — improves the capability for analyzing and interpreting these data via their inherent spatial characteristics.
The Geographical Edge 263
Point-of-Sale Data Capture and Customer Loyalty Programs
The widespread adoption of point-of-sale (POS) automation technology has given retailers the opportunity to improve nearly every aspect of their businesses. With electronic POS systems, detailed time, product, and price data are captured for every transaction, which has made planning, inventory management, buying, theft prevention, in-store promotion, and so on much more reliable. In addition, point-of-sale automation has opened the door to the development of individual consumer-based loyalty programs.
Modeled after frequent flyer programs offered by airlines, retailer loyalty programs confer such benefits as immediate cost savings, members-only deals, rebates at some threshold level of spending, redeemable points, and/or eligibility for drawings and contests, all to “reward” shoppers for giving up alternative shopping opportunities.
Schneiderman (1998) reports that nearly half of the U.S. population belongs to at least one loyalty program and that such programs are growing at a rate of approximately 11%
a year.
More importantly for researchers, most retail loyalty programs involve the use of specially coded credit/debit cards or other special scanner-readable cards, which contain consumer-specific identification information. When these cards are scanned at the point of purchase, data is captured which links the consumer to the time, day, products bought, prices, and so on. Analysis of these data over time can yield invaluable insights into consumer shopping processes, reactions to marketing efforts, and long-term patterns of behaviors at the individual as well as at the aggregate level. A variety of techniques being applied to this data include various forms of regression, factor analysis, cluster analysis, time series analysis, and chaid analysis. In addition, the fact that loyalty programs typically require members to provide name, address, and other relevant information about themselves gives researchers the opportunity to use an arsenal of geographic analysis tools to better understand the shopping and buying behaviors of current customers and to target new ones.
Geographic Information Systems
Geographic information systems (GIS) are the second technology bringing radical change to retail-oriented research. Prior to GIS, spatial data was difficult to obtain in a form necessary for meaningful marketing research (e.g., addresses) and expensive and subject to error when collected in a more analytically usable form (e.g., accurate relative distances). Awkward to handle and time-consuming to create, the number of paper maps needed to cover the market areas of a major U.S. retailer could reach the thousands. In addition, a map-centered approach did not lend itself to easy physical reproduction, and the analyses were more difficult to replicate or extend than non-spatial analysis results (e.g., a trade area map is specific to a particular store location; a regression model of consumers based on census data is not).
While clearly based on cartography (the science of mapmaking), GIS technology goes much farther than a computerized map. The capability of a GIS is significantly expanded
264 Allway, Murphy and Berkowitz
by combining the association of non-spatial descriptive data (attribute data) with spatial features in a visually interactive mode that supports changes in scale (e.g., zoom in/zoom out), the overlaying of different types of spatially-encoded information (i.e., like paper map overlays), and the creation and display of new geographic information (e.g., identifying the set of elements with certain attributes and within a specified distance of a geographic feature). Traditional (non-spatial) querying and analysis tools can be combined with spatial information once a geographic coordinate is associated with the item of interest (Murphy, 1995). For U.S. consumer data, the technique of geo-coding uses a pre-defined list of addresses and their spatial components along with a searching/
matching algorithm (Densham, 1991; Keenan, 1995).
Retailing was an early adopter of GIS, primarily for store location decisions (e.g., Baker
& Baker, 1993; Daniel, 1994; Foust & Botts, 1995). Once the store was located, however, the role of GIS often gave way to traditional analysis approaches (e.g., media-revenue recovery models) in which geography was a constant (e.g., the location of the store) or only slowly varying (e.g., a store’s trade area). With the increasing amount of customer- specific data being collected at the store level, however, the analysis of customers can increasingly exploit the spatial aspects of consumer behavior.
This Study
This chapter demonstrates some of the additional insights that a GIS-based analysis approach can offer in the study of the spatial diffusion in the context of a new loyalty card program. We show that the study of spatially-oriented consumer behaviors and the business strategies that result from analysis of these behaviors both benefit greatly from the application of GIS technology. The situation involves the launch and testing of a new loyalty card program by a very large U.S. retailer within a major metropolitan area.
This loyalty card program constituted a major effort on the part of the retailer to build store traffic, increase basket size, and increase shopping frequency while creating deeper relationship ties with its customer base. The large-scale launch effort for the loyalty card program included city-wide radio, a number of billboards, and professional in-store solicitation. Data capture was via checkout scanner, and every transaction in which the consumer “swiped” the card was recorded. According to company records, the launch of the program was highly successful, with an increase of nearly 30% in sales during the first few weeks of the program compared to the prior year.
Customer Loyalty POS Data
Detailed information on the full first year of the program was provided to the researchers, including the launch campaign, a name and address database of every cardholder, a purchasing occasion database, and a stock-keeping-unit (SKU) level sales database (both identified to the cardholder level) for three separate stores. When combined, the resulting data set covered well over one million distinct shopping trips and several million
The Geographical Edge 265
SKU-level product purchases. After narrowing the focus to consumers of a single representative store, and combining and organizing information, a data set was created which included a cardholder identifier, detailed street address, date of first card usage, date of last card usage (both recoded from day one through day 365), number of purchasing occasions during the year, total dollars spent in the store over the year (using the card), average amount spent per shopping visit, highest and lowest dollar amount spent on a shopping trip, duration of shopping activity (last day of card use minus first day) and shopping interval (average time between purchasing occasions).
Traditional Non-Spatial Analysis
While most analyses of customer loyalty programs do not take advantage of GIS technology, loyalty program data are valuable to retail decision-making. Sales data at the aggregate and at the individual levels can be tracked over time and patterns noted and modeled. Time series analysis, cluster analysis, logistic regression, and other tools can be used to visualize the timing of shopping, to distinguish between loyalty groups, and to estimate the impact of different marketing efforts on the loyalty base and the subgroups within it.
We first demonstrate a traditional diffusion-of-innovation approach to search for insights about the growth of and the prospects for the loyalty card population. Using Rogers’ (1962, 1983, 1995) and Mahajan, Muller, & Srivastava’s (1990) frameworks, each of the nearly 18,000 adopters was classified into an innovator, early adopter, early majority, late majority, or laggard group. Because these are assigned categories based on timing of the adoption relative to the pattern of overall adoptions, classification of particular individuals into these categories is accomplished by examining the temporal distribution of adoptions and looking for transition points following the percentage distribution guidelines of Rogers (1995).
Compared to some other innovations (telephone, automobiles, air conditioning), a loyalty card program has a short adoption cycle (less than 180 days versus decades), which is a factor in making the decisions about the cut-off between adopter groups. The classification that captured the dynamics of this data most accurately was to set the innovator cut-off after two days, which yielded 1,073 persons, or 6.0% of all eventual adopters. The early adopter stage of the process began on day three of the program and ran through day seven, when 18.5% of all eventual adopters had made their first purchase.
The cutoff for the early majority was made after the day 31 of the program, when 50.5%
of all eventual adopters had made their first purchase. The late majority category cutoff was made after the 120th day, with 84.7% of the total, while the last 15.3% of adopters were relegated to laggard status. The comparison of the percentages by adopter group for this study vs. Rogers (1995) is shown in Table 1.
As shown in Table 2, there are significant differences among the adoption groups in nearly every category of basic descriptor. This, in and of itself, is interesting and can lead to additional insights relevant to retailers. Such phenomena as cross-shopping, the number of new adopters as well as the number of deserters each week, the increase in new adoptions following a radio blitz or new round of promotion, increases or decreases in
266 Allway, Murphy and Berkowitz
overall card use, SKU’s bought, and customer loyalty can all be tracked without the use of geographical data.
Application of GIS
However, just the fact that address-specific information exists in the loyalty card database enables retailers to expand the value of this data many-fold. The insights available by the application of GIS technology to these data open the door to a level of analysis far beyond those of traditional retail researchers. To take advantage of the potential inherent in the spatial data captured by the customer loyalty/POS program, it was necessary to begin by preparing the customer data set for loading into a GIS program.
This involved the creation of a single data set from the three separate databases kept by the retailer — a customer ID-coded name and address database, a customer ID-coded purchase event database, and a customer ID-coded products purchased database.
Rogers (1995) This Study ADOPTER GROUP
Percent in Category
Cumulative Percent
Percent in Category
Cumulative Percent
Innovators 2.5 2.5 6.0 6.0
Early Adopters 13.5 16 13.5 18.5
Early Majority 34 50 32.0 50.5
Late Majority 34 84 34.2 84.7
Laggards 16 100 15.3 100
Table 1. Comparison of Adopter Group Classification
DIFFUSION STAGE Stage One Stage
Two Stage Three Total
ADOPTER GROUP Innovators Early Adopter
Early Majority
Late
Majority Laggards Total NUMBER IN GROUP 1,070 2,216 5,646 6,045 2,698 17,675
Pre - GIS Insights
Profile Characteristics Mean Mean Mean Mean Mean Mean Day of Adoption 1.47 4.85 18.40 68.62 167.29 55.58 Length of Loyalty Card Use (Days) 238.57 226.99 200.68 154.72 97.00 174.73 Interval Between Card Uses (Days) 23.23 25.17 28.05 29.26 22.35 26.94 Total Dollars on Card $772.90 $614.61 $426.85 $319.34 $240.22 $406.08 Number of Purchase Occasions 27.67 18.50 11.39 7.53 5.65 11.07 Dollars Spent per Trip $36.95 $38.99 $40.84 $47.39 $47.43 $43.62
Table 2. Statistical Profile of the Three Diffusion Stages and Five Adopter Groups: Pre- GIS Insights
The Geographical Edge 267
Together, these databases held over ten million lines of customer ID-coded information.
After combining by customer ID and isolating a single store’s activity, a data set of approximately 23,000 cardholding customers was produced.
Step 1: Geo-Coding & Creation of Distance Variables
The first task was to check and correct the coding of address characteristics and zip codes so geo-coding could proceed. The database was then loaded into a popular PC-based GIS. Using a built-in search algorithm, which matches addresses ranges with those in the national streets database, these loyalty program customers were geocoded to yield detailed eight-digit latitude and longitude figures (called lat-long) on each cardholder.
Approximately 15% percent of the addresses could not be matched, either because of address spelling errors, new construction (new streets not yet in the national street database), double-named streets, or colloquially named streets. These addresses were either hand-located or discarded, resulting in a final geo-coded database of 17,675 cardholders. Using similar address data, lat-longs were generated for the store, its competitors, and each billboard that advertised the loyalty program. Features of the GIS application were used to compute and add to the database additional spatial variables for each of the cardholder records including Euclidean distance from residence to the store, to the nearest billboard, and to each competitor, and the number of billboards and the number of competitors within 2.5 miles of the customer’s residence.
Finally, a “Neighborhood Interaction Field” (NIF) was created around each of the adopters of the loyalty card. After testing dozens of distance measures, a figure of .1 kilometers, or .06 miles was selected as the appropriate NIF radius around each adopter.
This distance covered approximately five to seven houses in all directions, more in tightly compressed housing configurations and fewer in areas with more distance between neighboring houses. Note that the spatial dispersion characteristics of other environ- ments (e.g., more urban or dense, more rural or distributed) can and should affect the radius chosen; the goal was to identify a practical measure to capture the likely residence- based communication influences on adoption effectively for the mid-western U.S.
suburban setting of the data set. All economic activity (previous adoptions, loyalty card- specific shopping behavior, spending, and so on by any of the other adopters) was captured for each cardholder and added to the data set as additional variables. None of these measures could have been generated without a GIS.
Step 2: Adding All Households in Market Area
A second data set of every household within a 35-mile radius of the store was created using geo-coded and mapped data from a direct mail list vendor. To truly understand the adoption process we need to study the innovation’s effect on not only the nearly 18,000 adopters but also on the approximately 300,000 households in the greater market area who did not become adopters of the loyalty program. The same distance-based measures were computed for non-adopters and added to the data set.