CHOICES™ on the Web

Burke Interactive programming and survey hosting makes it possible to use the Internet for studies employing Burke’s CHOICES™ discrete-choice approach.

Case History—Sprint’s Award-Winning Discrete Choice Research

Sprint recently was named an honorable mention winner in the 2001 EXPLOR awards for an online study employing CHOICES™.

The EXPLOR awards are designed to salute “Exemplary Performance and Leadership in Online Research” and are administered by the University of Wisconsin’s A.C. Nielsen Center for Marketing Research and the American Marketing Association. Stan Frear, group manager of Sprint’s consumer marketing research division, will discuss the study at the 2001 EXPLOR Forum, to be held November 15-17, 2001, at the Hotel Intercontinental in Chicago. Burke senior vice president Jeff Miller will also present an overview of online research at the event.

Study Overview

Technological advances, regulatory changes and corporate mergers are bringing with them a convergence of telecommunications and entertainment services. To help meet the challenges of this convergence, Sprint worked with Breakthru Strategies, an industry marketing consulting firm, to develop an overall approach to modeling the rapidly evolving market.

A discrete choice exercise using Burke’s proprietary CHOICES™ approach served as the foundation for much of the model. A total of 1,500 members of a U.S. online consumer panel were asked to select preferences from sets of individual and “bundled” services offered by a variety of national and regional providers of local, long-distance, and wireless telephone service, Internet access, and cable television and other entertainment content.

The survey was custom programmed by Burke Interactive so that service choices in each set were randomized within a complex set of constraints, resulting in data on 9,000 unique choice sets. Using a disaggregate discrete choice procedure enabled Sprint to predict what kinds of services individual consumers are interested in purchasing – and what competitive conditions make it likely that they will. Sprint has used the findings to make forecasts for specific communities with different demographic profiles and can also adapt the model as national demographic and technology-usage patterns change over the next several years.

More Information

·         Study Background: Telecommunications Convergence

·         The Target Population

·         The Choice Task

·         Product Attributes

·         Fully Randomized Design

·         Increasing Respondent Understanding

·         Analysis and Modeling


Study Background: Telecommunications Convergence

Telecommunications has seen a convergence of formerly disparate categories in recent years, with single providers now frequently offering a “bundle” of services. Such a bundle might include (in various combinations) local, long distance, and wireless telephone service, conventional and high-speed Internet access, and cable television or other entertainment content. Meeting the challenge of this convergence – brought on by technological advances, regulatory changes, and corporate mergers – is central to Sprint’s marketing efforts now and in the foreseeable future.

The landscape has been evolving so rapidly, in fact, that it has been difficult to predict what kinds of services consumers would be interested in purchasing and under what conditions. Sprint has conducted highly sophisticated research on consumer decision processes utilizing discrete choice modeling. This research has considered the choices consumers make in each of the converging telecommunications categories, as well as how these choices interact with each other when multiple services are offered by the same supplier. Data, however, were in some cases nearly obsolete before results were available and models could be built. The competitive landscape represented no longer existed, with new alternatives already available to consumers.

As a consequence, Sprint undertook the present project to build a more flexible model that could adapt to marketplace changes for at least five years into the future. Moving beyond just marketing research, this effort involved creating an entire framework for decision-making informed by multiple sources. This model has proved to be effective for year 2000 forecasting, and shows promise to be an adaptable and invaluable business tool for the future.

The Target Population

The predictive model was populated by data from a discrete choice exercise conducted by Burke Interactive among 1,500 members of a U.S. online consumer panel. Based on models from other research, Sprint was specifically interested in consumers falling in certain more technically-oriented segments of the population, which by definition overlap largely with current users of online services. Consequently, Internet-based data collection was very appropriate for reaching the Web-enabled target market of interest. An Internet panel was utilized as a sample source and a $10 incentive was paid.

The Choice Task

Each respondent viewed six different market scenarios (choice sets) in the discrete choice exercise.

Unlike many consumer choice models used in the telecommunications industry, this effort recognized that even as more services are marketed more frequently as bundles, many consumers will continue to opt to purchase them separately from different suppliers. Possible product categories therefore included six “individual” services (local telephone service, cable programming, Internet access, etc.), as well as four specific combinations of these six services likely to be sold as bundles (telephone service with Internet access, for instance).

Both of these alternatives (bundled or separate purchases) had to be included to produce a useful and comprehensive model of consumer behavior, even though doing so introduced complications into the data collection process. Not all product categories were available in each choice set, however, so that simulations could be conducted to model “before” and “after” the emergence of new products and estimate the impact of their introductions.

It was also deemed likely that some of the products would be offered by more than one supplier simultaneously, so within some choice sets, some products were available in more than one version. Within each choice set, there were actually between six and 13 product options.

The 10 product categories included some pairs which were mutually exclusive, such as cable television alone and cable television bundled with telephone service, and other pairs which were not mutually exclusive, such as residential telephone service and dial-up Internet access. Consequently, while respondents viewed themselves as choosing multiple services from a menu of up to 13 alternatives, they were actually choosing one of 79 possible combinations of services (all possible combinations with the mutually exclusive pairs not permitted).

Product Attributes

Each of the products was, in turn, defined by up to 11 attributes, including brand name, both recurring and non-recurring costs, and features such as included minutes of calling, upload and download speeds, and activation variables. Not all attributes applied to all product categories.

One challenge Burke Interactive faced in programming the study is that the largest providers of many telecommunications services are local companies with local brand names. To accurately capture the equity of these important local players, the brand names appearing in interview stimuli were customized to each local market. The interview covered 15 local markets, and each had its own unique providers (and different numbers of providers) of local telephone service, cable television, and Internet access.

Increasing the design complexity further were several prohibitions in the combinations of attribute levels that could appear in a given choice set. For instance, although Internet access upload and download speeds covered overlapping ranges, upload speed could be no faster than download speed. Certain speeds could only be combined with certain technologies (DSL vs. cable modem, for instance), and certain brands could only offer certain technologies (a cable company wouldn’t offer DSL). The ability to share “included calling minutes” across platforms (long distance with wireless, for instance) was contingent on both products being available from the same supplier and the product offering actually featuring some included minutes (yet another variable).

Fully Randomized Design

These contingencies, along with the overall size of the design and model, made a traditionally-constructed experimental design of a manageable size for the discrete choice exercise virtually impossible to develop. Instead of using such a static blocked design, one of the innovations of this research was employing a fully-randomized design.

Each choice set viewed by each respondent was created “on the fly” through random number generation, according to a highly complex algorithm adhering to a set of rules and constraints. With 1,500 respondents each viewing six choice sets, data were collected on a total of 9,000 unique choice sets.

Near perfect attribute level balance and orthogonality could be achieved where desired (each level appearing the same number of times and paired equally often with the levels of other attributes). Orthogonality was achieved not only between pairs of attributes within a single choice option, but also between pairs of attributes across different options. (In some cases, perfect balance was not desired, and it was a simple matter to make certain levels for which more data and precision was desired – “Sprint” as the level of “brand,” for instance – appear more frequently.)

No aliasing (confounding of main effects with high-order interactions between other variables) was required, so all interactions, even those of higher order and those not anticipated before data collection, could be estimated.

Increasing Respondent Understanding

Because of the complexity of the respondent choice task (up to 13 potentially unfamiliar options with up to 11 attributes each), several creative measures were taken to increase respondent understanding and thus the quality of the data. Some of these measures were enabled specifically by the online interviewing method.

Because some of the products and features were unfamiliar, respondents were exposed to full-color, annotated mock-ups of promotional materials – as realistic as possible with brand logos where appropriate – before beginning the discrete choice task.

Once in the choice task, respondents viewed the choice sets one at a time. Each set was presented as a grid, with names of various available options across the top and the attributes describing each in rows beneath them. Appearing at the top of the page were several “hyperlinks” which, when clicked, allowed the respondent to return and review different aspects of the background promotional materials in pop-up windows.

The left-to-right order in which the specific options were presented was varied to avoid any bias of order effects. This variation, however, was between respondents rather than within the choice sets seen by a particular respondent. For any one respondent, the order of the options was constant across the six choice sets he or she completed to keep the task less confusing.

Because not all options were available in all choice sets, a given option still did not always appear in exactly the same location for each respondent. Since visual space was at a premium and a goal was to minimize left-to-right scrolling by respondents, when an option was not available in a given choice set, blank columns were not left in the grid. Rather, as some options moved in or out of the set, others were naturally shifted to the left or right with a variable number of columns displayed.

To help respondents locate familiar products from set to set, color coding was used. The attribute text in the column describing each option was presented in a unique color that remained constant across choice sets. The color coding scheme was carried forward to the bottom of the page, where respondents actually clicked to select the services they would purchase, to help ensure that they were selecting the intended alternatives.

Analysis and Modeling

For the analysis, a 79-option discrete choice model was estimated (one option for each of the 79 possible combinations of services). Burke Interactive provided a stand-alone simulator based on the research sample. Because Burke’s proprietary CHOICES™ discrete-choice modeling was used to analyze consumer preferences at an individual level, choice tendencies could be presented as a function of a variety of demographic and current technology-usage variables.

The discrete choice exercise also served as the foundation for most of the parameters and effects in the larger flexible structural model developed by Breakthru Strategies. That framework also incorporates other estimates, especially of the impact of products and services not currently available. These products may be too foreign to consumers for them to provide reliable data about their behavior with regard to those options. As these products gain greater penetration, additional research will be used to refine the existing model, so that there will be no need to start over from scratch.

Of course, even among currently familiar products, prices and usage patterns will change, and consumers will over time assign different priorities to various options. The disaggregate nature of the model derived from the consumer discrete choice exercise allows key parameters to be readily updated. Because, for example, the relationship between demographics and preferences has been drawn, demographics themselves can be directly inputted into the model. The model can therefore adapt as national demographic and technology-usage patterns change over the next several years, and it can also be customized for specific communities with different demographic profiles. Such flexibility, for example, helped Sprint perform short-term market-specific planning in the fall of 2000.